49.6 Cognitive Processes 481 CogPrime's multi-memory design, it's natural to consider CogPrime's cognitive processes in terms of which memory subsystems they focus on (although, this is not a perfect mode of analysis, since some of the cognitive processes span multiple memory types). 49.6.1 Uncertain Logic for Declarative Knowledge One major decision made in the creation of CogPrime was that given the strengths and weak- nesses of current and near-future digital computers, uncertain logic is a good way to handle declarative knowledge. Of course this is not obvious nor is it the only possible route. Declarative knowledge can potentially be handled in other ways; e.g. in a hierarchical network architecture, one can make declarative knowledge emerge automatically from procedural and sensorimotor knowledge, as is the goal in the Numenta and DeSTIN designs reviewed in Chapter 4 of Part 1. It seems clear that the human brain doesn't contain anything closely parallel to formal logic - even though one can ground logic operations in neural-net dynamics as explored in Chapter 34, this sort of grounding leads to "uncertain logic enmeshed with a host of other cognitive dynamics" rather than "uncertain logic as a cleanly separable cognitive process." But contemporary digital computers are not brains - they lack the human brain's capacity for cheap massive parallelism, but have a capability for single-operation speed and precision far exceeding the brain's. In this way computers and formal logic are a natural match (a fact that's not surprising given that Boolean logic lies at the foundation of digital computer opera- tions). Using uncertain logic is a sort of compromise between brainlike messiness and fuzziness, and computerlike precision. An alternative to using uncertain logic is using crisp logic and in- corporating uncertainty as content within the knowledge base - this is what SOAR does, for example, and it's not a wholly unworkable approach. But given that the vast mass of knowledge needed for confronting everyday human reality is highly uncertain, and that this knowledge of- ten needs to be manipulated efficiently in real-time, it seems to us there is a strong argument for embedding uncertainty in the logic. Many approaches to uncertain logic exist in the literature, including probabilistic and fuzzy approaches, and one conclusion we reached in formulating CogPrime is that none of them was adequate on its own — leading us, for example, to the conclusion that to deal with the problems facing a human-level AG!, an uncertain logic must integrate imprecise probability and fuzziness with a broad scope of logical constructs. The arguments that both fuzziness and probability are needed scent hard to counter - these two notions of uncertainty are qualitatively different yet both appear cognitively necessary. The argument for using probability in an AGI system is assailed by some AGI researchers such as Pei Wang, but we are swayed by the theoretical arguments in favor of probability theory's mathematically fundamental nature, as well as the massive demonstrated success of probability theory in various areas of narrow AI and applied science. However, we are also swayed by the arguments of Pei Wang, Peter %Valley and others that using single-number probabilities to represent truth values leads to untoward complexities related to the tabulation and manipulation of amounts of evidence. This has led us to an imprecise probability based approach; and then technical arguments regarding the limitations of standard imprecise probability formalisms has led us to develop our own "indefinite probabilities" formalism. The PLN logic framework is one way of integrating imprecise probability and fuzziness in a logical formalism that encompasses a broad scope of logical constructs. It integrates term logic EFTA00624628

482 49 Summary of Argument for the CogPrime Approach and predicate logic - a feature that we consider not necessary, but very convenient, for AGI. Either predicate or term logic on its own would suffice, but each is awkward in certain cases, and integrating them as done in PLN seems to result in more elegant handling of real-world inference scenarios. Finally, PLN also integrates intensional inference in an elegant manner that demonstrates integrative intelligence - it defines intension using pattern theory, which binds inference to pattern recognition and hence to other cognitive processes in a conceptually appropriate way. Clearly PLN is not the only possible logical formalism capable of serving a human-level AGI system; however, we know of no other existing, fleshed-out formalism capable of fitting the bill. In part this is because PLN has been developed as part of an integrative AGI project whereas other logical formalisms have mainly been developed for other purposes, or purely theoretically. Via using PLN to control virtual agents, and integrating PLN with other cognitive processes, we have tweaked and expanded the PLN formalism to serve all the roles required of the "declarative cognition" component of an AGI system with reasonable elegance and effectiveness. 49.6.2 Program Learning for Procedural Knowledge Even more so than declarative knowledge, procedural knowledge is represented in many different ways in the Al literature. The human brain also apparently uses multiple mechanisms to embody different kinds of procedures. So the choice of how to represent procedures in an AGI system is not particularly obvious. However, there is one particular representation of procedures that is particularly well-suited for current computer systems, and particularly well-tested in this context: programs. In designing CogPrime, we have acted based on the understanding that programs are a good way to represent procedures - including both cognitive and physical-action procedures, but perhaps not including low-level motor-control procedures. Of course, this begs the question of programs in what programming language, and in this context we have made a fairly traditional choice, using a special language called Combo that is essentially a minor variant of LISP, and supplying Combo with a set of customized primitives intended to reduce the length of the typical programs CogPrime needs to learn and use. What differentiates this use of LISP from many traditional uses of LISP in Al is that we are only using the LISP-ish representational style for procedural knowledge, rather than trying to use it for everything. One test of whether the use of Combo programs to represent procedural knowledge makes sense is whether the procedures useful for a CogPrime system in everyday human environments have short Combo representations. We have worked with Combo enough to validate that they generally do in the virtual world environment - and also in the physical-world environment if lower-level motor procedures are supplied as primitives. That is, we are not convinced that Combo is a good representation for the procedure a robot needs to do to move its fingers to pick up a cup, coordinating its movements with its visual perceptions. It's certainly possible to represent this sort of thing in Combo, but Combo may be an awkward tool. However, if one represents low-level procedures like this using another method, e.g. learned cell assemblies in a hierarchical network like DeSTIN, then it's very feasible to make Combo programs that invoke these low-level procedures, and encode higher-level actions like "pick up the cup in front of you slowly and quietly, then hand it to Jim who is standing next to you." EFTA00624629

49.6 Cognitive Processes 483 Having committed to use programs to represent many procedures, the next question is how to learn programs. One key conclusion we have come to via our empirical work in this area is that some form of powerful program normalization is essential. Without normalization, it's too hard for existing learning algorithms to generalize from known, tested programs and draw useful uncertain conclusions about untested ones. We have worked extensively with a generalization of Holman's "Elegant Normal Form" in this regard. For learning normalized programs, we have come to the following conclusions: • for relatively straightforward procedure learning problems, hilldimbing with random restart and a strong Occam bias is an effective method • for more difficult problems that elude hinclimbing, probabilistic evolutionary program learn- Mg is an effective method The probabilistic evolutionary program learning method we have worked with most in OpenCog is MOSES, and significant evidence has been gathered showing it to be dramatically more effective than genetic programming on relevant classes of problems. However, more work needs to be done to evaluate its progress on complex and difficult procedure learning problems. Alternate, related probabilistic evolutionary program learning algorithms such as PLEASURE have also been considered and may be implemented and tested as well. 49.6.5 Attention Allocation There is significant evidence that the brain uses some sort of "activation spreading" type method to allocate attention, and many algorithms in this spirit have been implemented and utilized in the Al literature. So, we find ourselves in agreement with many others that activation spreading is a reasonable way to handle attentional knowledge (though other approaches, with greater overhead cost, may provide better accuracy and may be appropriate in some situations). We also agree with many others who have chosen Hebbian learning as one route of learning associative relationships, with more sophisticated methods such as information-geometric ones potentially also playing a role. Where CogPrime differs from standard practice is in the use of an economic metaphor to reg- ulate activation spreading. In this matter CogPrime is broadly in agreement with Eric Baum's arguments about the value of economic methods in Al, although our specific use of economic methods is very different from his. Baum's work (e.g. Hayek [13an0 ID embodies more complex and computationally expensive uses of artificial economics, whereas we believe that in the con- text of a neural-symbolic network, artificial economics is an effective approach to activation spreading; and CogPrime's ECAN framework seeks to embody this idea. ECAN can also make use of more sophisticated and expensive uses of artificial currency when large amount of system resources are involved in a single choice, rendering the cost appropriate. One major choice made in the CogPrime design is to focus on two kinds of attention: proces- sor (represented by ShortTermImportance) and memory (represented by LongTermImportance). This is a direct reflection of one of the key differences between the von Neumann architecture and the human brain: in the former but not the latter, there is a strict separation between mem- ory and processing in the underlying compute fabric. We carefully considered the possibility of using a larger variety of attention values, and in Chapter 23 we presented some mathematics and concepts that could be used in this regard, but for reasons of simplicity and computational EFTA00624630

484 49 Summary of Argument for the CogPrime Approach efficiency we are currently using only STI and LTI in our OpenCogPrime implementation, with the passibility of extending further if experimentation proves it necessary. 49.6.4 Internal Simulation and Episodic Knowledge For episodic knowledge, as with declarative and procedural knowledge, CogPrime has opted for a solution motivated by the particular strengths of contemporary digital computers. When the human brain runs through a "mental movie" of past experiences, it doesn't do any kind of accurate physical simulation of these experiences. But that's not because the brain wouldn't benefit from such - it's because the brain doesn't know how to do that sort of thing! On the other hand, any modern laptop can run a reasonable Newtonian physics simulation of everyday events, and more fundamentally can recall and manage the relative positions and movements of items in an internal 3D landscape paralleling remembered or imagined real-world events. With this in mind, we believe that in an AGI context, simulation is a good way to handle episodic knowledge; and running an internal "world simulation engine" is an effective way to handle simulation. CogPrime can work with many different simulation engines; and since simulation technology is continually advancing independently of AGI technology, this is an area where AGI can buy some progressive advancement for free as time goes on. The subtle issues here regard interfacing between the simulation engine and the rest of the mind: mining meaningful information out of simulations using pattern mining algorithms; and more subtly, figuring out what simulations to run at what times in order to answer the questions most relevant to the AGI system in the context of achieving its goals. We believe we have architected these interactions in a viable way in the CogPrime design, but we have tested our ideas in this regard only in some fairly simple contexts regarding virtual pets in a virtual world, and much more remains to be done here. 49.6.5 Low-Level Perception and Action The centrality or otherwise of low-level perception and action in human intelligence is a matter of ongoing debate in the AI community. Sonic feel that the essence of intelligence lies in cognition and/or language, with perception and action having the status of "peripheral devices." Others feel that modeling the physical world and one's actions in it is the essence of intelligence, with cognition and language emerging as side-effects of these more fundamental capabilities. The CogPrime architecture doesn't need to take sides in this debate. Currently we are experimenting both in virtual worlds, and with real-world robot control. The value added by robotic versus virtual embodiment can thus be explored via experiment rather than theory, and may reveal nuances that no one currently foresees. As noted above, we are tmconfident of CogPrime's generic procedure learning or pattern recognition algorithms in terms of their capabilities to handle large amounts of raw sensorimotor data in real time, and so for robotic applications we advocate hybridizing CogPrime with a separate (but closely cross-linked) system better customized for this sort of data, in line with our general hypothesis that Hybridization of one's integrative neural-symbolic system with a spatiotemporally hierarchical deep learning system is an effective way to handle representation EFTA00624631

49.7 Fulfilling the "Cognitive Equation" 485 and learning of low-level sensorimotor knowledge. While this general principle doesn't depend on any particular approach, DeSTIN is one example of a deep learning system of this nature that can be effective in this context We have not yet done any sophisticated experiments in this regard - our current experiments using OpenCog to control robots involve cruder integration of OpenCog with perceptual and motor subsystems, rather than the tight hybridization described in Chapter 26. Creating such a hybrid system is last" a matter of software engineering, but testing such a system may lead to many surprises! 49.6.6 Goals Given that we have characterized general intelligence as "the ability to achieve complex goals in complex environments," it should be plain that goals play a central role in our work. However, we have chosen not to create a separate subsystem for intentional knowledge, and instead have concluded that one effective way to handle goals is to represent them declaratively, and allocate attention among them economically. An advantage of this approach is that it automatically provides integration between the goal system and the declarative and attentional knowledge systems. Goals and subgoaLs are related using logical links as interpreted and manipulated by PLN, and attention is allocated among goals using the STI dynamics of ECAN, and a specialized variant based on RFS's (requests for service). Thus the mechanics of goal management is handled using uncertain inference and artificial economics, whereas the figuring-out of how to achieve goals is done integratively, relying heavily on procedural and episodic knowledge as well as PLN and ECAN. The combination of ECAN and PLN seems to overcome the well-known shortcomings found with purely neural-net or purely inferential approaches to goals. Neural net approaches gener- ally have trouble with abstraction, whereas logical approaches are generally poor at real-time responsiveness and at tuning their details quantitatively based on experience. At least in prin- ciple, our hybrid approach overcomes all these shortcomings; though of current, it has been tested only in fairly simple cases in the virtual world. 49.7 Fulfilling the "Cognitive Equation" A key claim based on the notion of the "Cognitive Equation" posited in Chaotic Logic [Coe9lJ is that it is important for an intelligent system to have some way of recognizing large-scale patterns in itself, and then embodying these patterns as new, localized knowledge items in its memory. This dynamic introduces a feedback dynamic between emergent pattern and substrate, which is hypothesized to be critical to general intelligence under feasible computational resources. It also ties in nicely with the notion of "glocal memory" - essentially positing a localization of some global memories, which naturally will result in the formation of some glocal memories. One of the key ideas underlying the CogPrime design is that given the use of a neural-symbolic network for knowledge representation, a graph-mining based "map formation" heuristic is one good way to do this. EFTA00624632

486 49 Summary of Argument for the CogPrime Approach Map formation seeks to fulfill the Cognitive Equation quite directly, probably more directly than happens in the brain. Rather than relying on other cognitive processes to implicitly recog- nize overall system patterns and embody them in the system as localized memories (though this implicit recognition may also happen), the MapFormation MindAgent explicitly carries out this process. Mostly this is done using fairly crude greedy pattern mining heuristics, though if really subtle and important patterns seem to be there, more sophisticated methods like evolutionary pattern mining may also be invoked. It seems possible that this sort of explicit approach could be less efficient than purely implicit approaches; but, there is no evidence for this, and it may actually provide increased efficiency. And in the context of the overall CogPrime design, the explicit NIapFormation approach seems most natural. 49.8 Occam's Razor The key role of "Occam's Razor" or the urge for simplicity in intelligence has been observed by many before (going back at least to Occam himself, and probably earlier!), and is fully embraced in the CogPrime design. Our theoretical analysis of intelligence, presented in Chapter 2 of Part 1 and elsewhere, portrays intelligence as closely tied to the creation of procedures that achieve goals in environments in the simplest possible way. And this quest for simplicity is present in many places throughout the CogPrime design, for instance • In MOSES and hilIclimbing, where program compactness is an explicit component of pro- gram tree fitness • In PLN, where the backward and forward chainers. explicitly favor shorter proof chains, and intensional inference explicitly characterizes entities in terms of their patterns (where patterns are defined as compact characterizations) • In pattern mining heuristics, which search for compact characterizations of data • In the forgetting mechanism, which seeks the smallest set of Atoms that will allow the regeneration of a larger set of useful Atoms via modestly-expensive application of cognitive processes • Via the encapsulation of procedural and declarative knowledge in simulations, which in many cases provide a vastly compacted form of storing real-world experiences Like cognitive synergy and emergent networks, Occam's Razor is not something that is imple- mented in a single place in the CogPrime design, but rather an overall design principle that underlies nearly every part of the system. 49.8.1 Mind Geometry The three mind-geometric principles outlined in Appendix ?? are: • syntax-semantics correlation • cognitive geometrodynamics • cognitive synergy EFTA00624633

49.8 Occam's Razor 487 The key role of syntax-semantics correlation in CogPrime is clear. It plays an explicit role in MOSES. In PLN, it is critical to inference control, to the extent that inference control is based on the extraction of patterns from previous inferences. The syntactic structures are the inference trees, and the semantic structures are the inferential conclusions produced by the trees. History-guided inference control assumes that prior similar trees will be a good starting-point for getting results similar to prior ones - i.e. it assumes a reasonable degree of syntax-semantics correlation. Also, without a correlation between the core elements used to generate an episode, and the whole episode, it would be infeasible to use historical data mining to understand what core elements to use to generate a new episode - and creation of compact, easily manipulable seeds for generating episodes would not be feasible. Cognitive geometrodynamics is about finding the shortest path from the current state to a goal state, where distance is judged by an appropriate metric including various aspects of computational effort. The ECAN and effort management frameworks attempt to enforce this, via minimizing the amount of effort spent by the system in getting to a certain conclusion. MindAgents operating primarily on one kind of knowledge (e.g. MOSES, PLN) may for a time seek to follow the shortest paths within their particular corresponding memory spaces; but then when they operate more interactively and synergetically, it becomes a matter of finding short paths in the composite mindspace corresponding to the combination of the various memory types. Finally, cognitive synergy is thoroughly and subtly interwoven throughout CogPrime. In a way the whole design is about cognitive synergy - it's critical for the design's functionality that it's important that the cognitive processes associated with different kinds of memory can appeal to each other for assistance in overcoming bottlenecks in a manner that: a) works in "real time'; i.e. on the time scale of the cognitive processes internal processes; b) enables each cognitive process to act in a manner that is sensitive to the particularities of each others' internal representations. Recapitulating in a bit more depth, recall that another useful way to formulate cognitive synergy as follows. Each of the key learning mechanisms underlying CogPrime is susceptible to combinatorial explosions. As the problems they confront become larger and larger, the per- formance gets worse and worse at an exponential rate, because the number of combinations of items that mast be considered to solve the problems grows exponentially with the problem size. This could be viewed as a deficiency of the fundamental design, but we don't view it that way. Our view is that combinatorial explosion is intrinsic to intelligence. The task at hand is to dampen it sufficiently that realistically large problems can be solved, rather than to eliminate it entirely. One possible way to dampen it would be to design a single, really clever learning algorithm - one that was still susceptible to an exponential increase in computational require- ments as problem size increases, but with a surprisingly small exponent. Another approach is the mirrorhouse approach: Design a bunch of learning algorithms, each focusing on different aspects of the learning process, and design them so that they each help to dampen each others' combinatorial explosions. This is the approach taken within CogPrime. The component algo- rithms are clever on their own - they are less susceptible to combinatorial explosion than many competing approaches in the narrow-AI literature. But the real meat of the design lies in the intended interactions between the components, manifesting cognitive synergy. EFTA00624634

488 49 Summary of Argument for the CogPrime Approach 49.9 Cognitive Synergy To understand more specifically how cognitive synergy works in CogPrime, in the following sub- sections we will review some synergies related to the key components of CogPrime as discussed above. These synergies are absolutely critical to the proposed functionality of the CogPrime system. Without them, the cognitive mechanisms are not going to work adequately well, but are rather going to succumb to combinatorial explosions. The other aspects of CogPrime - the cognitive architecture, the knowledge representation, the embodiment framework and associ- ated developmental teaching methodology - are all critical as well, but none of these will yield the critical emergence of intelligence without cognitive mechanisms that effectively scale. And, in the absence of cognitive mechanisms that effectively scale on their own, we mast rely on cognitive mechanisms that effectively help each other to scale. The reasons why we believe these synergies will exist are essentially qualitative: we have not proved theorems regarded these syn- ergies, and we have observed them in practice only in simple cases so far. However, we do have some ideas regarding how to potentially prove theorems related to these synergies, and some of these are described in Appendix H. 49.9.1 Synergies that Help Inference The combinatorial explosion in PLN is obvious: forward and backward chaining inference are both fundamentally explosive processes, reined in only by pruning heuristics. This means that for nontrivial complex inferences to occur, one needs really, really clever pruning heuristics. The CogPrime design combines simple heuristics with pattern mining, MOSES and economic attention allocation as pruning heuristics. Economic attention allocation assigns importance levels to Atoms, which helps guide pruning. Greedy pattern mining is used to search for patterns in the stored corpus of inference trees, to see if there are any that can be used as analogies for the current inference. And MOSES comes in when there is not enough information (from importance levels or prior inference history) to make a choice, yet exploring a wide variety of available options is unrealistic. In this case, MOSES tasks may be launched, pertinently to the leaves at the fringe of the inference tree, under consideration for expansion. For instance, suppose there is an Atom A at the fringe of the inference tree, and its importance hasn't been assessed with high confidence, but a number of items B are known so that: MemberLink A B Then, MOSES may be used to learn various relationships characterizing A, based on recognizing patterns across the set of B that are suspected to be members of A. These relationships may then be used to assess the importance of A more confidently, or perhaps to enable the inference tree to match one of the patterns identified by pattern mining on the inference tree corpus. For example, if MOSES figures out that: SimilarityLink G A then it may happen that substituting G in place of A in the inference tree, results in something that pattern mining can identify as being a good (or poor) direction for inference. EFTA00624635

49.10 Synergies that Help MOSES 489 49.10 Synergies that Help MOSES MOSES's combinatorial explosion is obvious: the number of possible programs of size N increases very rapidly with N. The only way to get around this is to utilize prior knowledge, and as much as possible of it. When solving a particular problem, the search for new solutions must make use of prior candidate solutions evaluated for that problem, and also prior candidate solutions (including successful and unsuccessful ones) evaluated for other related problems. But, extrapolation of this kind is in essence a contextual analogical inference problem. In some cases it can be solved via fairly straightforward pattern mining; but in subtler cases it will require inference of the type provided by PLN. Also, attention allocation plays a role in figuring out, for a given problem A, which problems B are likely to have the property that candidate solutions for B are useful information when looking for better solutions for A. 49.10.1 Synergies that Help Attention Allocation Economic attention allocation, without help from other cognitive processes, is just a very sim- ple process analogous to "activation spreading" and "Hebbian learning" in a neural network. The other cognitive processes are the things that allow it to more sensitively understand the attentional relationships between different knowledge items (e.g. which sorts of items are often usefully thought about in the same context, and in which order). 49.10.2 Further Synergies Related to Pattern Mining Statistical, greedy pattern mining is a simple process, but it nevertheless can be biased in various ways by other, more subtle processes. For instance, if one has learned a population of programs via MOSES, addressing some particular fitness function, then one can study which items tend to be utilized in the same programs in this population. One may then direct pattern mining to find patterns combining these items found to be in the MOSES population. And conversely, relationships denoted by pattern mining may be used to probabilistically bias the models used within MOSES. Statistical pattern mining may also help PLN by supplying it with information to work on. For instance, conjunctive pattern mining finds conjunctions of items, which may then be combined with each other using PLN, leading to the formation of more complex predicates. These conjunctions may also be fed to MOSES as part of an initial population for solving a relevant problem. Finally, the main interaction between pattern mining and MOSES/PLN is that the former may recognize patterns in links created by the latter. These patterns may then be fed back into MOSES and PLN as data. This virtuous cycle allows pattern mining and the other, more expensive cognitive processes to guide each other. Attention allocation also gets into the game, by guiding statistical pattern mining and telling it which terms (and which combinations) to spend more time on. EFTA00624636

490 49 Summary of Argument for the CogPrime Approach 49.10.3 Synergies Related to Map Formation The essential synergy regarding map formation is obvious: Maps are formed based on the HebbianLinks created via PLN and simpler attentional dynamics, which are based on which Atoms are usefully used together, which is based on the dynamics of the cognitive processes doing the "using." On the other hand, once maps are formed and encapsulated, they feed into these other cognitive processes. This synergy in particular is critical to the emergence of self and attention. What has to happen, for map formation to work well, is that the cognitive processes must utilize encapsulated maps in a way that gives rise overall to relatively clear clusters in the network of HebbianLinks. This will happen if the encapsulated maps are not too complex for the system's other learning operations to understand. So, there must be useful coordinated attentional patterns whose corresponding encapsulated-map Atoms are not too complicated. This has to do with the system's overall parameter settings, but largely with the settings of the attention allocation component. For instance. this is closely tied in with the limited size of "attentional focus" (the famous 7 +/- 2 number associated with humans' and other mammals short term memory capacity). If only a small number of Atoms are typically very important at a given point in time, then the maps formed by grouping together all simultaneously highly important things will be relatively small predicates, which will be easily reasoned about - thus keeping the "virtuous cycle" of map formation and comprehension going effectively. 49.11 Emergent Structures and Dynamics We have spent much more time in this book on the engineering of cognitive processes and structures, than on the cognitive processes and structures that must emerge in an intelligent system for it to display human-level AGI. However, this focus should not be taken to represent a lack of appreciation for the importance of emergence. Rather, it represents a practical focus: engineering is what we must do to create a software system potentially capable of AGI, and emergence is then what happens inside the engineered AGI to allow it to achieve intelligence. Emergence must however be taken carefully into account when deciding what to engineer! One of the guiding ideas underlying the CogPrime design is that an AGI system with ade- quate mechanisms for handling the key types of knowledge mentioned above, and the capability to explicitly recognize large-scale pattern in itself should upon sustained interaction with an appropriate environment in pursuit of appropriate goals, emerge a variety of com- plex structures in its internal knowledge network, including (but not limited to): a hierarchical network, representing both a spatiotemporal hierarchy and an approximate "default inheritance" hierarchy, cross-linked; a heterarchical network of associativity, roughly aligned with the hierar- chical network; a self network which is an approximate micro image of the whole network; and inter-reflecting networks modeling self and others, reflecting a "mirrorhouse" design pattern. The dependence of these posited emergences on the environment and goals of the AGI system should not be underestimated. For instance, PLN and pattern mining don't have to lead to a hierarchical structured Atomspace, but if the AGI system is placed in an environment which is itself hierarchically structured, then they very likely will do so. And if this environment consists of hierarchically structured language and culture, then what one has is a system of minds with hierarchical networks, each reinforcing the hierarchality of each others' networks. Similarly, EFTA00624637

49.12 Ethical AC! 491 integrated cognition doesn't have to lead to mirrorhouse structures, but integrated cognition about situations involving other minds studying and predicting and judging each other, is very likely to do so. What is needed for appropriate emergent structures to arise in a mind, is mainly that the knowledge representation is sufficiently flexible to allow these structures, and the cognitive processes are sufficiently intelligent to observe these structures in the environment and then minor them internally. Of course, it also doesn't hurt if the internal structures and processes are at least slightly biased toward the origination of the particular high-level emergent structures that are characteristic of the system's environment/goals; and this is indeed the case with CogPrime - biases toward hierarchical, heterarchical, dual and mirrorhouse networks are woven throughout the system design, in a thoroughgoing though not extremely systematic way. 49.12 Ethical AGI Creating an AGI with guaranteeably ethical behavior seems an infeasible task; but of course, no human is guaranteeably ethical either, and in fact it seems almost guaranteed that in any moderately large group of humans there are going to be some with strong propensities for extremely unethical behaviors, according to any of the standard human ethical codes. One of our motivations in developing CogPrime has been the belief that an AC! system, if supplied with a commonsensically ethical goal system and an intentional component based on rigorous uncertain inference, should be able to reliably achieve a much higher level of commonsensically ethical behavior than any human being. Our explorations in the detailed design of CogPrime's goal system have done nothing to degrade this belief. While we have not yet developed any CogPrime system to the point where experimenting with its ethics is meaningful, based on our understanding of the current design it seems to us that • a typical CogPrime system will display a much more consistent and less conflicted and confused motivational system than any human being, due to its explicit orientation toward carrying out actions that (based on its knowledge) rationally seem most likely to lead to achievement of its goals • if a CogPrime system is given goals that are consistent with commonsensical human ethics (say, articulated in natural language), and then educated in an ethics-friendly environment such as a virtual or physical school, then it is reasonable to expect the CogPrime system will ultimately develop an advanced (human adult level or beyond) form of commmonsensical human ethics Human ethics is itself wracked with inconsistencies, so one cannot expect a rationality-based AGI system to precisely mirror the ethics of any particular human individual or cultural system. But given the degree to which general intelligence represents adaptation to its environment, and interpretation of natural language depends on life history and context, it seems very likely to us that a CogPrime system. if supplied with a human-commonsense-ethics based goal system and then raised by compassionate and intelligent humans in a school-type environment, would arrive at its own variant of human-commonsense-ethics. The AGI system's ethics would then interact with human ethical systems in complex ways, leading to ongoing evolution of both systems and the development of new cultural and ethical patterns. Predicting the future is EFTA00624638

492 49 Summary of Argument for the CogPrime Approach difficult even in the absence of radical advanced technologies, but our intuition is that this path has the potential to lead to beneficial outcomes for both human and machine intelligence. 49.13 Toward Superhuman General Intelligence Human-level AGI is a difficult goal, relative to the current state of scientific understanding and engineering capability, and most of this book has been focused on our ideas about how to achieve it. However, we also suspect the CogPrime architecture has the ultimate potential to push beyond the human level in many ways. As part of this suspicion we advance the claim that once sufficiently advanced, a CogPrime system should be able to radically self-improve via a variety of methods, including supercompilation and automated theorem-proving. Supercompilation allows procedures to be automatically replaced with equivalent but mas- sively more time-efficient procedures. This is particularly valuable in that it allows AI algorithms to learn new procedures without much heed to their efficiency, since supercompilation can al- ways improve the efficiency afterwards. So it is a real boon to automated program learning. Theorem-proving is difficult for current narrow-Al systems, but for an AGI system with a deep understanding of the context in which each theorem exists, it should be much easier than for human mathematicians. So we envision that ultimately an AGI system will be able to design itself new algorithms and data structures via proving theorems about which ones will best help it achieve its goals in which situations, based on mathematical models of itself and its environment. Once this stage is achieved, it seems that machine intelligence may begin to vastly outdo human intelligence, leading in directions we cannot now envision. While such projections may seem science-fictional, we note that the CogPrime architecture explicitly supports such steps. If human-level AGI is achieved within the CogPrime framework, it seems quite feasible that profoundly self-modifying behavior could be achieved fairly shortly thereafter. For instance, one could take a human-level CogPrime system and teach it computer science and mathematics, so that it fully understood the reasoning underlying its own design, and the whole mathematics curriculum leading up the the algorithms underpinning its cognitive processes. 49.13.1 Conclusion What we have sought to do in these pages is, mainly, • to articulate a theoretical perspective on general intelligence, according to which the cre- ation of a human-level AGI doesn't require anything that extraordinary, but "merely" an appropriate combination of closely interoperating algorithms operating on an appropriate multi-type memory system, utilized to enable a system in an appropriate body and envi- ronment to figure out how to achieve its given goals • to describe a software design (CogPrime ) that, according to this somewhat mundane but theoretically quite well grounded vision of general intelligence, appears likely (according to a combination of rigorous and heuristic arguments) to be able to lead to human-level AGI using feasible computational resources EFTA00624639

49.13 Toward Superhuman General Intelligence 493 • to describe some of the preliminary lessons we've learned via implementing and experiment- ing with aspects of the CogPrime design, in the OpenCog system In this concluding chapter we have focused on the "combination of rigorous and heuristic argu- ments" that lead us to consider it likely that CogPrime has the potential to lead to human-level AGI using feasible computational resources. We also wish to stress that not all of our arguments and ideas need to be 100% correct in order for the project to succeed. The quest to create AGI Ls a mix of theory, engineering, and scientific and unscientific experimentation. If the current CogPrime design turns out to have significant shortcomings, yet still brings us a significant percentage of the way toward human-level AGI, the results obtained along the path will very likely give us clues about how to tweak the design to more effectively get the rest of the way there. And the OpenCog platform is extremely flexible and extensible, rather than being tied to the particular details of the CogPrime design. While we do have faith that the CogPrime design as described here has human-level AGI potential, we are also pleased to have a development strategy and implementation platform that will allow us to modify and improve the design in whatever suggestions are made by our ongoing experimentation. Many great achievements in history have seemed more magical before their first achievement than afterwards. Powered flight and spaceflight are the most obvious examples, but there are many others such as mobile telephony, prosthetic limbs, electronically deliverable books, robotic factory workers, and so on. We now even have wireless transmission of power (one can recharge cellphones via wifi), though not yet as ambitiously as Tesla envisioned. We very strongly suspect that human-level AGI Ls in the same category as these various examples: an exciting and amazing achievement, which however is achievable via systematic and careful application of fairly mundane principles. We believe computationally feasible human-level intelligence is both complicated (involving many interoperating parts, each sophisticated in their own right) and complex (in the sense of involving many emergent dynamics and structures whose details are not easily predictable based on the parts of the system) ... but that neither the complication nor the complexity is an obstacle to engineering human-level AGI. Furthermore, while ethical behavior is a complex and subtle matter for humans or machines, we believe that the production of human-level AGIs that are not only intelligent but also ben- eficial to humans and other biological sentiences, is something that is probably tractable to achieve based on a combination of careful AGI design and proper AGI education and "parent- ing." One of the motivations underlying our design has been to create an artificial mind that has broadly humanlike intelligence, yet has a more rational and self-controllable motivational system than humans, thus ultimately having the potential for a greater-than-human degree of ethical reliability alongside its greater-than-human intelligence. In our view, what is needed to create human-level AGI is not a new scientific breakthrough, nor a miracle, but "merely" a sustained effort over a number of years by a moderate-sized team of appropriately-trained professionals, completing the implementation of the design in this book and then parenting and educating the resulting implemented system. CogPrime is by no means the only possible path to human-level AGI, but we believe it is considerably more fully thought-through and fleshed-out than any available alternatives. Actually, we would love to see CogPrime and a dozen alternatives simultaneously pursued - this may seem ambitious, but it would cost a fraction of the money currently spent on other sorts of science or engineering, let alone the money spent on warfare or decorative luxury items. We strongly suspect that, in hindsight, our human and digital descendants will feel amazed that their predecessors allocated EFTA00624640

494 49 Summary of Argument for the CogPrime Approach so few financial and attentional resources to the creation of powerful AGI, and consequently took so long to achieve such a fundamentally straightforward thing. EFTA00624641

Chapter 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment 50.1 Introduction AGI design necessarily leads one into some rather abstract spaces — but being a human-like intelligence in the everyday world Ls a pretty concrete thing. If the CogPrime research program is successful, it will result not just in abstract ideas and equations, but rather in real AGI robots carrying out tasks in the world, and AGI agents in virtual worlds and online digital spaces conducting important business, doing science, entertaining and being entertained by us, and so forth. With this in mind, in this final chapter we will bring the discussion closer to the concrete and everyday, and pursue a thought experiment of the form "How would a completed CogPrime system carry, out this specific task?" The task we will use for this thought-experiment is one we have used as a running example now and then in the preceding chapters. We consider the case of a robotically or virtually embodied CogPrime system, operating in a preschool type environment, interacting with a human whom it already knows and given the task of "Build me something with blocks that I haven't seen before." This target task is fairly simple, but it is complex enough to involve essentially every one of CogPrime's processes, interacting in a unified way. It involves simple, grounded creativity of the sort that normal human children display every day - and which, we conjecture, is structurally and dynamically basically the same as the creativity underlying the genius of adult human creators like Einstein, Dali, Dostoevsky, Hendrix, and so forth ... and as the creativity that will power massively capable genius machines in future. We will consider the case of a simple interaction based on the above task where: 1. The human teacher tells the CogPrime agent "Build me something with blocks that I haven't seen before." 2. After a few false starts, the agent builds something it thinks is appropriate and says 'Do you like it?" 3. The human teacher says "It's beautiful. What is it?" 4. The agent says "It's a car man" land indeed. the construct has 4 wheels and a chassis vaguely like a car, but also a torso, arms and head vaguely like a person] Of course, a complex system like CogPrime could carry out an interaction like this internally in many different ways, and what is roughly described here is just one among many possibilities. 495 EFTA00624642

496 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment First we will enumerate a number of CogPrime processes and explain some ways that each one may help CogPrime carry out the target task. Then we will give a more evocative narrative, conveying the dynamics that would occur in CogPrime while carrying out the target task, and mentioning how each of the enumerated cognitive processes as it arises in the narrative. 50.2 Roles of Selected Cognitive Processes Now we review a number of the more interesting CogPrime cognitive processes mentioned in previous chapters of the hook, for each one indicating one or more of the roles it might play in helping a CogPrime system carry out the target task. Note that this list is incomplete in many senses, e.g. it doesn't list all the cognitive processes, nor all the roles played by the ones listed. The purpose is to give an evocative sense of the roles played by the different parts of the design in carrying out the task. • Chapter 19 (OpenCog Fnunework) - Freezing/defrosting. • When the agent builds a structure from blocks and decides it's not good enough to show off to the teacher, what does it do with the detailed ideas and thought process underlying the structure it built? If it doesn't like the structure so much, it may just leave this to the generic forgetting process. But if it likes the structure a lot, it may want to increase the VLTI (Very Long Term Importance) of the Atoms related to the structure in question, to be sure that these are stored on disk or other long-term storage, even after they're deemed sufficiently irrelevant to be pushed out of RAM by the forgetting mechanism. • When given the target task, the agent may decide to revive from disk the mind- states it went through when building crowd-pleasing structures from blocks before, so as to provide it with guidance. • Chapter 22 (Emotion, Motivation, Attention and Control) - Cognitive cycle. • While building with blocks, the agent's cognitive cycle will be dominated by per- ceiving, acting on, and thinking about the blocks it is building with. • When interacting with the teacher, then interaction-relevant linguistic, perceptual and gestural processes will also enter into the cognitive cycle. - Emotion. The agent's emotions will fluctuate naturally as it carries out the task. • If it has a goal of pleasing the teacher, then it will experience happiness as its expectation of pleasing the teacher increases. • If it has a goal of experiencing novelty, then it will experience happiness as it creates structures that are novel in its experience. • If it has a goal of learning, then it will experience happiness as it learns new things about blocks construction. • On the other hand, it will experience unhappiness as its experienced or predicted satisfaction of these goals decreases. - Action selection EFTA00624643

50.2 Roles of Selected Cognitive Processes 497 In dialoguing with the teacher, action selection will select one or more DialogueCon- troller schema to control the conversational interaction (based on which DC schema have proved most effective in prior similar situations. When the agent wants to know the teacher's opinion of its construct, what is happening internally is that the "please teacher" Goal Atom gets a link of the conceptual form (Implication "find out teacher's opinion of my current con- struct" "please teacher"). This link may be created by PLN inference, prob- ably largely by analogy to previously encountered similar situations. Then, Goallmportance is spread from the "please teacher" Goal Atom to the "find out teacher's opinion of my current construct" Atom (via the mechanism of sending an RFS package to the latter Atom). More inference causes a link (Implication "ask the teacher for their opinion of my current construct" "find out teacher's opinion of my current construct") to be formed, and the "ask the teacher for their opinion of my current construct" Atom to get Goallmportance also. Then PredicateSchematization causes the predicate "ask the teacher for their opinion of my current construct" to get turned into an actionable schema, which gets Goallmportance, and which gets pushed into the ActiveSchemaPool via Goal- driven action selection. Once the schema version of "ask the teacher for their opinion of my current construct" is in the ActiveSchemaPool, it then invokes natural language generation Tasks, which lead to the formulation of an English sentence such as "Do you like it?" When the teacher asks "It's beautiful. What is it?", then the NL comprehension MindAgent identifies this as a question, and the "please teacher" Goal Atom gets a link of the conceptual form (Implication "answer the question the teacher just asked" "please teacher"). This follows simply from the knowledge ( Implication ("teacher has just asked a question" AND "I answer the teacher's question") ("please teacher")), or else from more complex knowledge refining this Impli- cation. From this point, things proceed much as in the case "Do you like it?" described just above. Consider a schema such as "pick up a red cube and place it on top of the long red block currently at the top of the structure" (let's call this P). Once P is placed in the ActiveSchemaPool, then it runs and generates more specific procedures, such as the ones needed to find a red cube, to move the agent's arm toward the red cube and grasp it, etc. But the execution of these specific low-level procedures is done via the ExecutionManager, analogously to the execution of the specifics of generating a natural language sentence from a collection of semantic relationships. Loosely speaking, reaching for the red cube and turning simple relationships into a simple sentences, are considered as "automated processes" not requiring holistic engagement of the agent's mind. What the generic, more holistic Action Selection mechanism does in the present context is to figure out to put P in the ActiveSchemaPool in the first place. This occurs because of a chain such as: P predictively implies (with a certain probabilistic weight) "completion of the car-man structure", which in turn predictively implies "completion of a structure that is novel to the teacher," which in turn predictively implies "please the teacher," which in turn implies "please others," which is assumed an Ubergoal (a top-level system goal). EFTA00624644

498 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment - Goal Atoms. As the above items make clear, the scenario in question requires the initial Goal Atoms to be specialized, via the creation of more and more particular subgoals suiting the situation at hand. - Context Atoms. • Knowledge of the context the agent is in can help it disambiguate language it hears, e.g. knowing the context is blocks-building helps it understand which sense of the word "blocks" is meant. • On the other hand, if the context is that the teacher is in a bad mood, then the agent might know via experience that in this context, the strength of (Implication "ask the teacher for their opinion of my current construct" "find out teacher's opinion of my current construct") is lower than in other contexts. - Context formation. • A context like blocks-building or teacher in a bad mood" may be formed by cluster- ing over multiple experience-sets, i.e. forming Atoms that refer to spatiotemporally grouped sets of percepts/concepts/actions, and grouping together similar Atoms of this nature into clusters. • The Atom referring to the cluster of experience-sets involving blocks-building will then survive a S an Atom if it gets involved in relationships that are important or have surprising truth values. If many relationships have significantly different truth- value inside the blocks-building context than outside it, this means it's likely that the blocks-building ConceptNode will remain as an Atom with reasonably high LTI, so it can be used as a context in future. - Time-dependence of goals. Many of the agent's goals in this scenario have different importances over different time scales. For instance "please the teacher" is important on multiple time-scales: the agent wants to please the teacher in the near term but also in the longer term. But a goal like "answer the question the teacher just asked" has an intrinsic time-scale to it; if it's not fulfilled fairly rapidly then its importance goes away. • Chapter 23 (Attention allocation) - ShortTermlmportance versus LongTermlmportance. While conversing, the con- cepts and immediately involved in the conversation (including the Atoms describing the agents in the conversation) have very high STI. While building, Atoms representing to the blocks and related ideas about the structures being built (e.g. images of cars and people perceived or imagined in the past) have very high STI. But the reason these Atoms are in RAM prior to having their STI boosted due to their involvement in the agent's activities, is because they had their LTI boosted at some point in the past. And after these Atoms leave the AttentionalFocus and their STI reduces, they will have boosted LTI and hence likely remain in RAM for a long while, to be involved in "background thought", and in case they're useful in the AttentionalFocus again. - HebbianLink formation. As a single example, the car-man has both wheels and arms, so now a Hebbian association between wheels and arms will exist in the agent's memory, to potentially pop up again and guide future thinking. The very idea of a car-man likely emerged partly due to previously formed HebbianLinks - because people were often seen sitting in cars, the association between person and car existed, which made the car concept and the human concept natural candidates for blending. - Data mining the System Activity Table. The HebbianLinks mentioned above may have been formed via mining the SystemActivityTable EFTA00624645

50.2 Roles of Selected Cognitive Processes 499 - ECAN based associative memory. When the agent thinks about making a car, this spreads importance to various Atoms related to the car concept, and one thing this does is lead to the emergence of the car attractor into the AttentionalFocus. The different aspects of a car are represented by heavily interlinked Atoms, so that when some of them become important, there's a strong tendency for the others to also become important - and for "car" to then emerge as an attractor of importance dynamics. - Schema credit assignment. • Suppose the agent has a subgoal of placing a certain blue block on top of a certain red block. It may use a particular motor schema for carrying out this action - involving, for instance, holding the blue block above the red block and then gradually lowering it. If this schema results in success (rather than in, say, knocking down the red block), then it should get rewarded via having its STI and LTI boosted and also having the strength of the link between it and the subgoal increased. • Next, suppose that a certain cognitive schema (say, the schema of running multiple related simulations and averaging the results, to estimate the success probability of a motor procedure) was used to arrive at the motor schema in question. Then this cognitive schema may get passed some importance from the motor schema, and it will get the strength of its link to the goal increased. In this way credit passes backwards from the goal to the various schema directly or indirectly involved in fulfilling it. - Forgetting. If the agent builds many structures from blocks during its lifespan, it will accumulate a large amount of perceptual memory. • Chapter 24 (Goal and Action Selection). Much of the use of the material in this chapter was covered above in the bullet point for Chapter 22, but a few more notes are: - Transfer of RFS between goals. Above it was noted that the link (Implication "ask the teacher for their opinion of my current construct" "find out teacher's opinion of my current construct") might be formed and used as a channel for Goallmportance spreading. - Schema Activation. Supposing the agent is building a man-car, it may have car- building schema and man-building schema in its ActiveSchemaPool at the same time, and it may enact both of them in an interleaved manner. But if each tend to require two hands for their real-time enaction, then schema activation will have to pass back and forth between the two of them, so that at any one time, one is active whereas the other one is sitting in the ActiveSchemaPool waiting to get activated. - Goal Based Schema Learning. To take a fairly low-level example, suppose the agent has the (sub)goal of making an arm for a blocks-based person (or man-car), given the presence of a blocks-based torso. Suppose it finds a long block that seems suitable to be an arm. It then has the problem of figuring out how to attach the arm to the body. It may try out several procedures in its internal simulation world, until it finds one that works: hold the arm in the right position white one end of it rests on top of some block that is part of the torso, then place some other block on top of that end, then slightly release the arm and see if it falls. If it doesn't fall, leave it. If it seems about to fall, then place something heavier atop it, or shove it further in toward the center of the torso. The procedure learning process could be MOSES here, or it could be PLN. • Chapter 25 (Procedure Evaluation) EFTA00624646

500 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment - Inference Based Procedure Evaluation. A procedure for man-building such as "first put up feet, then put up legs, then put up torso, then put up arms and head" may be synthesized from logical knowledge (via predicate schematization) but without filling in the details of how to carry out the individual steps, such as "put up legs." If a procedure with abstract (ungrounded) schema like PutUpTorso is chosen for execu- tion and placed into the ActiveSchemaPool, then in the course of execution, inferential procedure evaluation must be used to figure out how to make the abstract schema ac- tionable. The CoalDrivenActionSelection MindAgent must make the choice whether to put a not-fully-grounded schema into the ActiveSchemaPool, rather than grounding it first and then making it active; this is the sort of choice that may be made effectively via learned cognitive schema. • Chapter 26 (Perception and Action) - ExperienceDB. No person remembers every blocks structure they ever saw or built, except maybe some autists. But a CogPrime can store all this information fairly easily, in its ExperienceDB, even if it doesn't keep it all in RAM in its AtomSpace. It can also store everything anyone ever said about blocks structures in its vicinity. - Perceptual Pattern Mining. - Object Recognition. Recognizing structures made of blocks as cars, people, houses, etc. requires fairly abstract object recognition, involving identifying the key shapes and features involved in an object-type, rather than just going by simple visual similarity. - Hierarchical Perception Networks. If the room is well-lit, it's easy to visually iden- tify individual blocks within a blocks structure. If the room is darker, then more top- down processing may be needed - identifying the overall shape of the blocks structure may guide one in making out the individual blocks. - Hierarchical Action Networks. Top-down action processing tells the agent that, if it wants to pick up a block, it should move its arm in such a way as to get its hand near the block, and then move its hand. But if it's still learning how to do that sort of motion, more likely it will do this, but then start moving its its hand and find that it's hard to get a grip on the block - and then have to go back and move its arm a little differently. Iterating between broader arm/hand movements and more fine-grained hand/finger movements is an instance of information iteratively passing up and down a hierarchical action network. - Coupling of Perception and Action Networks. Picking up a block in the dark is a perfect example of rich coupling of perception and action networks. Feeling the block with the fingers helps with identifying blocks that can't be clearly seen. • Chapter 30 (Procedure Learning) - Specification Based Procedure Learning. • Suppose the agent has never seen a horse, but the teacher builds a number of blocks structures and calls them horses, and draws a number of pictures and calls them horses. This may cause a procedure learning problem to be spawned, where the fitness function is accuracy at distinguishing horses from non-horses. • Learning to pick up a block is specification-based procedure learning, where the specification is to pick up the block and grip it and move it without knocking down the other stuff near the block. - Representation Building. EFTA00624647

50.2 Roles of Selected Cognitive Processes 501 • In the midst of building a procedure to recognize horses, MOSES would experi- mentally vary program nodes recognizing visual features into other program nodes recognizing other visual features • In the midst of building a procedure to pick up blocks, MOSES would experimentally vary program nodes representing physical movements into other nodes representing physical movements • In both of these cases, MOSES would also carry out the standard experimen- tal variations of mathematical and control operators according to its standard representation-building framework • Chapter 31 (Imitative, Reinforcement and Corrective Learning) - Reinforcement Learning. • Motor procedures for placing blocks (in simulations or reality) will get rewarded if they don't result in the blocks structure falling down, punished otherwise. • Procedures leading to the teacher being pleased, in internal simulations (or in re- peated trials of scenarios like the one under consideration), will get rewarded; pro- cedures leading to the teacher being displeased will get punished. - Imitation Learning. If the agent has seen others build with blocks before, it may summon these memories and then imitate the actions it has seen others take. - Corrective Learning. This would occur if the teacher intervened in the agent's block- building and guided him physically - e.g. steadying his shaky arm to prevent him from knocking the blocks structure over. • Chapter 32 (Hillclimbing) - Complexity Penalty. In learning procedures for manipulating blocks, the complexity penalty will militate against procedures that contain extraneous steps. • Chapter 33 (Probabilistic Evolutionary• Procedure Learning) - Supplying Evolutionary Learning with Long-Term Memory. Suppose the agent has previously built people from clay, but never from blocks. It may then have learned a "classification model" predicting which clay people will look appealing to humans, and which won't. It may then transfer this knowledge, using PLN, to form a classification model predicting which blocks-people will look appealing to humans, and which won't. - Fitness Function Estimation via Integrative Intelligence. To estimate the fitness of a procedure for, say, putting an arm on a blocks-built human, the agent may try out the procedure in the internal simulation world; or it may use PLN inference to reason by analogy to prior physical situations it's observed. These allow fitness to be estimated without actually trying out the procedure in the environment. • Chapter 34 (Probabilistic Logic Networks) - Deduction. This is a tall skinny structure; tall skinny structures fall down easily; thus this structure may fall down easily. - Induction. This teacher is talkative; this teacher is friendly; therefore the talkative are generally friendly. - Abduction. This structure has a head and arms and torso; a person has a head and arms and torso; therefore this structure is a person. EFTA00624648

502 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment - PLN forward chaining. What properties might a car-man have, based on inference from the properties of cars and the properties of men? - PLN backward chaining. • An inference target might be: Find X so that X looks something like a wheel and can be attached to this Mocks-chassis, and I can find four fairly similar copies. • Or: Find the truth value of the proposition that this structure looks like a car. - Indefinite truth values. Consider the deductive inference 'This is a tall skinny struc- ture; tall skinny structures fall down easily; thus this structure may fall down easily." In this case, the confidence of the second premise may be greater than the confidence of the first premise, which may result in an intermediate confidence for the conclusion, according to the propagation of indefinite probabilities through the PLN deduction rule. - Intensional inference. Is the blocks-structure a person? According to the definition of intensional inheritance, it shares many informative properties with people (e.g. having arms, torso and head), so to a significant extent, it is a person. - Confidence decay. The agent's confidence in propositions regarding building things with blocks should remain nearly constant. The agent's confidence in propositions re- garding the teacher's taste should decay more rapidly. This should occur because the agent should observe that, in general, propositions regarding physical object manipula- tion tend to retain fairly constant truth value, whereas propositions regarding human tastes tend to have more rapidly decaying truth value. • Chapter 35 (Spatiotemporal Inference) - Temporal reasoning. Suppose, after the teacher asks 'What is it?", the agent needs to think a while to figure out a good answer. But maybe the agent knows that it's rude to pause too long before answering something to a direct question. Temporal reasoning helps figure out "how long is too long" to wait before answering. - Spatial reasoning. Suppose the agent puts shoes on the wheels of the car. This is a joke relying on the understanding that wheels hold a car up, whereas feet hold a person up, and the structure is a car-man. But it also relies on the spatial inferences that: the car's wheels are in the right position for the man's feet (below the torso); and, the wheels are below the car's chassis just like a person's feet are below its torso. • Chapter 36 (Inference Control) - Evaluator Choice as a Bandit Problem. In doing inference regarding how to make a suitably humanlike arm for the blocks-man, there may be a choice between multiple inference pathways, perhaps one that relies on analogy to other situations building arms, versus one that relies on more general reasoning about lengths and weights of blocks. The choice between these two pathways will be made randomly with a certain probabilistic bias assigned to each one, via prior experience. - Inference Pattern Mining. The probabilities used in choosing which inference path to take. are determined in part by prior experience - e.g. maybe it's the case that in prior situations of building complex blocks structures, analogy, has proved a better guide than naive physics, thus the prior probability of the analogy inference pathway will be nudged up. - PLN and Bayes Nets. What's the probability that the blocks-man's hat will fall off if the man-car is pushed a little bit to simulate driving? This question could be resolved in many ways (e.g. by internal simulation), but one possibility is inference. If EFTA00624649

50.2 Roles of Selected Cognitive Processes 503 this is resolved by inference, it's the sort of conditional probability calculation that could potentially be done faster if a lot of the probabilistic knowledge from the AtomSpace were summarized in a Bayes Net. Updating the Bayes net structure can be slow, so this is probably not appropriate for knowledge that is rapidly shifting; but knowledge about properties of blocks structures may be fairly persistent after the agent has gained a fair bit of knowledge by playing with blocks a lot. • Chapter 37 (Pattern Mining) - Greedy Pattern Mining. • "Push a tall structure of blocks and it tends to fall down" is the sort of repetitive pattern that could easily be extracted from a historical record of perceptions and (the agent's and others') actions via simple greedy pattern mining algorithm. • If there is a block that is shaped like a baby's rattle, with a long slender handle and then a circular shape at the end, then greedy pattern mining may be helpful due to having recognized the pattern that structures like this are sometimes rattles - and also that structures like this are often stuck together, with the handle part connected sturdily to the circular part. - Evolutionary Pattern Mining. 'Push a tall structure of blocks with a wide base and a gradual narrowing toward the top and it may not fall too badly" is a more complex pattern that may not be found via greedy mining, unless the agent has dealt with a lot of pyramids. • Chapter 38 (Concept Formation) - Formal Concept Analysis. Suppose there are many long, slender blocks of different colors and different shapes (some cylindrical, some purely rectangular for example). Learning this sort of concept based on common features is exactly what FCA is good at (and when the features are defined fuzzily or probahilistically, it's exactly what uncertain FCA is good at). Learning the property of "slender" itself is another example of something uncertain FCA is good at - it would learn this if there were many concepts that preferentially involved slender things (even though formed on the basis of concepts other than slenderness) - Conceptual Blending. The concept of a "car-man" or "man-car" is an obvious instance of conceptual blending. The agents know that building a man won't surprise the teacher, and nor will building a car ... but both "man" and "car" may pop to the forefront of its mind (i.e. get a briefly high STI) when it thinks about what to build. But since it knows it has to do something new or surprising, there may be a cognitive schema that boosts the amount of funds to the ConceptBlending MindAgent, causing it to be extra-active. In any event, the ConceptBlending agent seeks to find ways to combine important concepts; and then PLN explores these to see which ones may be able to achieve the given goal of surprising the teacher (which includes subgoals such as actually being buildable). • Chapter 39 (Dimensional Embedding) - Dimensional Embedding. When the agent needs to search its memory for a previ- ously seen blocks structure similar to the currently observed one - or for a previously articulated thought similar to the one it's currently trying to articulate - then it needs to to a search through its large memory for "an entity similar to X" (where X is a EFTA00624650

504 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment structure or a thought). This kind of search can be quite computationally difficult - but if the entities in question have been projected into an embedding space, then it's quite rapid. (The cost is shifted to the continual maintenance of the embedding space, and its periodic updating; and there is some error incurred in the projection, but in many cases this error is not a show-stopper.) Embedding Based Inference Control. Rapid search for answers to similarity or in- heritance queries can be key for guiding inference in appropriate directions; for instance reasoning about how to build a structure with certain properties, can benefit greatly from rapid search for previously-encountered substructures currently structurally or functionally similar to the substructures one desires to build. • Chapter 40 (Simulation and Episodic Memory) - Fitness Estimation via Simulation . One way to estimate whether a certain blocks structure is likely to fall down or not, is to build it in one's "mind's eye" and see if the physics engine in one's mind's-eye causes it to fall down. This is something that in many cases will work better for CogPrime than for humans, because CogPrime has a more mathematically accurate physics engine than the human mind does; however, in cases that rely heavily on naive physics rather than, say, direct applications of Newton's Laws, then CogPrime's simulation engine may tmderperform the typical human mind. - Concept Formation via Simulation . Objects may be joined into categories using uncertain FCA, based on features that they are identified to have via "simulation exper- iments" rather than physical world observations. For instance, it may be observed that pyramid-shaped structures fall less easily than pencil-shaped tower structures - and the concepts corresponding to these two categories may be formed - from experiments run in the internal simulation world, perhaps inspired by isolated observations in the physical world. - Episodic Memory. Previous situations in which the agent has seen similar structures built, or been given similar problems to solve, may be brought to mind as "episodic movies" playing in the agent's memory. By watching what happens in these replayed episodic movies, the agent may learn new declarative or procedural knowledge about what to do. For example, maybe there was some situation in the agent's past where it saw someone asked to do something surprising, and that someone created something funny. This might (via a simple PLN step) bias the agent to create something now, which it has reason to suspect will cause others to laugh. • Chapter 41 (Integrative Procedure Learning) - Concept-Driven Procedure Learning. Learning the concept of "horse", as discussed above in the context of Chapter 30, is an example of this. - Predicate Schematization. The synthesis of a schema for man-building, as discussed above in the context of Chapter 25, is an example of this. • Chapter 42 (Map Formation) - Map Formation. The notion of a car involves many aspects: the physical appearance of cars, the way people get in and out of cars, the ways cars drive, the noises they make, etc. All these aspects are represented by Atoms that are part of the car map, and are richly interconnected via HebbianLinks as well as other links. EFTA00624651

50.2 Roles of Selected Cognitive Processes 505 - Map Encapsulation . The car map forms implicitly via the interaction of multiple cognitive dynamics, especially ECAN. But then the MapEncapstdation MindAgent may do its pattern mining and recognize this map explicitly, and form a PredicateNode encapsulating it. This PredicateNode may then be used in PLN inference, conceptual blending, and so forth (e.g. helping with the formation of a concept like car-man via blending). • Chapter 44 (Natural Language Comprehension) - Experience Based Diszunbiguation. The particular dialogue involved in the present example doesn't require any nontrivial word sense disambiguation. But it does require parse selection, and semantic interpretation selection: In "Build me something with blocks," the agent has no trouble understanding that "blocks" means "toy building blocks" rather than, say, "city blocks", based on many possible mechanisms, but most simply importance spreading. "Build me something with blocks" has at least three interpretations: the building could be carried out using blocks with a tool; or the thing built could be presented alongside blocks; or the thing built could be composed of blocks. The latter is the most commonsensical interpretation for most humans, but that is because we have heard the phrase "building with blocks" used in a similarly grounded way before (as well as other similar phrases such as "playing with Legos", etc., whose meaning helps militate toward the right interpretation via PLN inference and importance spreading). So here we have a simple example of experience-based disambiguation, where experiences at various distances of association from the current one are used to help select the correct parse. A subtler form of semantic disambiguation is involved in interpreting the clause "that I haven't seen before." A literal-minded interpretation would say that this requirement is fulfilled by any blocks construction that's not precisely identical to one the teacher has seen before. But of course, any sensible human knows this is an idiomatic clause that means "significantly different from anything I've seen before." This could be determined by the CogPrime agent if it has heard the idiomatic clause before, or if it's heard a similar idiomatic phrase such as "something I've never done before." Or, even if the agent has never heard such an idiom before, it could potentially figure out the intended meaning simply because the literal-minded interpretation would be a pointless thing for the teacher to say. So if it knows the teacher usually doesn't add useless modificatory clauses onto their statements, then potentially the agent could guess the correct meaning of the phrase. • Chapter 46 (Language Generation) - Experience-Based Knowledge Selection for Language Generation. When the teacher asks ' hat is it?", the agent must decide what sort of answer to give. Within the confines of the QuestionAnswering DialogueController, the agent could answer "A structure of blocks", or "A part of the physical world", or "A thing", or "Mine." (Or, if it were running another DC, it could answer more broadly, e.g. "None of your business," etc.). However, the QA DC tells it that, in the present context, the most likely desired answer is one that the teacher doesn't already know; and the most important property of the structure that the teacher doesn't obviously already know is the fact that it EFTA00624652

506 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment depicts a "car man." Also, memory, of prior conversations may bring up statements like 'It's a horse" in reference to a horse built of blocks, or a drawing of a horse, etc. - Experience-Based Guidance of Word and Syntax Choice . The choice of phrase "car man" requires some choices to be made. The agent could just as well say "It's a man with a car for feet" or "It's a car with a human upper body and head" or "It's a car centaur," etc. A bias toward simple expressions would lead to "car man." If the teacher were known to prefer complex expressions, then the agent might be biased toward expressing the idea in a different way. • Chapter 48 (Natural Language Dialogue) - Adaptation of Dialogue Controllers. The QuestionAsking and QuestionAnswer- ing DialogueControllers both get reinforcement from this interaction, for the specific internal rules that led to the given statements being made. 50.3 A Semi-Narrative Treatment Now we describe how a CogPriine system might carry out the specified task in a semi-narrative form, weaving in the material from the previous section as we go along, and making some more basic points as well. The semi-narrative covers most but not all of the bullet points from the previous section, but with some of the technical details removed; and it introduces a handful of new examples not given in the bullet points. The reason this is called a semi-narrative rather than a narrative is that there is no particular linear order to the processes occurring in each phase of the situation described here. CogPrime's internal cognitive processes do not occur in a linear narrative; rather, what we have is a complex network of interlocking events. But still, describing some of these events concretely in a manner correlated with the different stages of a simple interaction, may have some expository value. The human teacher tells the CogPrime agent "Build me something with blocks that I haven't seen before." Upon hearing this, the agent's cognitive cycles are dominated by language processing and retrieval from episodic and sensory memory. The agent may decide to revive from disk the mind-states it went through when building human-pleasing structures from blocks before, so as to provide it with guidance It will likely experience the emotion of happiness, because it anticipates the pleasure of getting rewarded for the task in future. The ubergoal of pleasing the teacher gets active (gets funded significantly with STI currency), as it becomes apparent there are fairly clear ways of fulfilling that goal (via the subgoal S of building blocks structures that will get positive response from the teacher). Other ubergoals like gaining knowledge are not funded as much with STI currency just now, as they are not immediately relevant. Action selection, based on ImplicationLinks derived via PLN (between various possible activ- ities and the subgoal S) causes it to start experimentally building some blocks structures. Past experience with building (turned into ImplicationLinks via mining the SystemActivityTable) tells it that it may want to build a little bit in its internal simulation world before building in the external world, causing STI currently to flow to the simulation MindAgent. EFTA00624653

50.3 A Semi-Narrative Treatment 507 The Atom corresponding to the context blocks-building gets high STI and is pushed into the AttentionalFocus, making it likely that many future inferences will occur in this context. Other Atoms related to this one also get high STI (the ones in the blocks-building map, and others that are especially related to blocks-building in this particular context). After a few false starts, the agent builds something it thinks is appropriate and says "Do you like it?" Now that the agent has decided what to do to fulfill its well-funded goal, its cognitive cycles are dominated by action, perception and related memory access and concept creation. An obvious subgoal is spawned: build a new structure now, and make this particular structure under construction appealing and novel to the teacher. This subgoal has a shorter time scale than the high level goal. The subgoal gets some currency from its supergoal using the mechanism of RFS spreading. Action selection must tell it when to continue building the same structure and when to try a new one, as well as more micro level choices. Atoms related to the currently pursued blocks structure get high STI. After a failed structure (a "false start") is disassembled, the corresponding Atoms lose STI dramatically (leaving AF) but may still have significant LTI, so they can be recalled later as appropriate. They may also have VLTI so they will be saved to disk later on if other things push them out of RAM due to getting higher LTI. Meanwhile everything that's experienced from the external world goes into the Experi- enceDB. Atoms representing different parts of aspects of the same blocks structure will get Hebbian links between them, which will guide future reasoning and importance spreading. Importance spreading helps the system go from an idea for something to build (say, a rock or a car) to the specific plans and ideas about how to build it, via increasing the STI of the Atoms that will be involved in these plan and ideas. If something apparently good is done in building a blocks structure, then other processes and actions that helped lead to or support that good thing, get passed some STI from the Atoms representing the good thing, and also may get linked to the Goal Atom representing "good" in this context. This leads to reinforcement learning. The agent may play with building structures and then seeing what they most look like, thus exercising abstract object recognition (that uses procedures learned by MOSES or hillclimbing, or uncertain relations learned by inference, to guess what object category a given observed collection of percepts mast likely falls into). Since the agent has been asked to come up with something surprising, it knows it should probably try to formulate some new concepts - because it has learned in the past, via Sys- temActivityTable mining, that often newly formed concepts are surprising to others. So, more STI currency is given to concept formation MindAgents, such as the ConceptualBlending Mind Agent (which, along with a lot of stuff that gets thrown out or stored for later use, comes up with "car-man"). When the notion of "car" is brought to mind, the distributed map of nodes corresponding to "car" get high STI. When car-man is formed, it is reasoned about (producing new Atoms), but it also serves as a nexus of importance-spreading, causing the creation of a distributed car-man map. If the goal of making an arm for a man-car occurs, then goal-driven schema learning may be done to learn a procedure for arm-making (where the actual learning is done by MOSES or EFTA00624654

508 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment If the agent Ls building a man-car, it may have man-building and car-building schema in its ActiveSchemaPool at the same time, and SchemaActivation may spread back and forth between the different modules of these two schema. If the agent wants to build a horse, but has never seen a horse made of blocks (only various pictures and movies of horses), it may uses MOSES or hillclimbing internally to solve the problem of creating a horse-recognizer or a horse-generator which embodies appropriate abstract properties of horses. Here as in all cases of procedure learning, a complexity penalty rewards simpler programs, from among all programs that approximately fulfill the goals of the learning process. If a procedure being executed has some abstract parts, then these may be executed by inferential procedure evaluation (which makes the abstract parts concrete on the fly in the course of execution). To guess the fitness of a procedure for doing something (say; building an arm or recognizing a horse), inference or simulation may be used, as well as direct evaluation in the world. Deductive, inductive and abductive PLN inference may be used in figuring out what a blocks structure will look or act like like before building it (it's tall and thin so it may fall down; it won't be bilaterally symmetric so it won't look much like a person; etc.) Backward-chaining inference control will help figure out how to assemble something matching a certain specification e.g. how to build a chassis based on knowledge of what a chassis looks like. Forward chaining inference (critically including intensional relationships) will be used to estimate the properties that the teacher will perceive a given specific structure to have. Spatial and temporal algebra will be used extensively in this reasoning, within the PLN framework. Coordinating different parts of the body - say an arm and a hand - will involve importance spreading (both up and down) within the hierarchical action network, and from this network to the hierarchical perception network and the heterarchical cognitive network. In looking up Atoms in the AtomSpace, sonic have truth values whose confidences have decayed significantly (e.g. those regarding the teacher's tastes), whereas others have confidences that have hardly decayed at all (e.g. those regarding general physical properties of blocks). Finding previous blocks structures similar to the current one (useful for guiding building by analogy to past experience) may be done rapidly by searching the system's internal dimensional- embedding space. As the building process occurs, patterns mined via past experience (tall things often fall down) are used within various cognitive processes (reasoning, procedure learning, concept cre- ation, etc.); and new pattern mining also occurs based on the new observations made as different structures are build and experimented with and destroyed. Simulation of teacher reactions, based on inference from prior examples, helps with the evaluation of possible structures, and also of procedures for creating structures. As the agent does all this, it experiences the emotion of curiosity (likely among other emo- tions), because as it builds each new structure it has questions about what it will look like and how the teacher would react to it. The human teacher says "It's beautiful. What is it?" The agent says "It's a car man" Now that the building is done and the teacher says something, the agent's cognitive cycles are dominated by language understanding and generation. The Atom representing the context of talking to the teacher gets high STI, and is used as the context for many ensuing inferences. EFTA00624655

50.4 Conclusion 509 Comprehension of "it" uses anaphor resolution based on a combination of ECAN and PLN inference based on a combination of previously interpreted language and observation of the external world situation. The agent experiences the emotion of happiness because the teacher has called its creation beautiful, which is recognizes as a positive evaluation - so the agent knows one of its ubergoals ("please the teacher") has been significantly fulfilled. The goal of pleasing the teacher causes the system to want to answer the question. So the QuestionAnswering DialogueController schema gets paid a lot and gets put into the Ac- tiveSchemaPool. In reaction to the question asked, this DC chooses a semantic graph to speak, then invokes NL generation to say it. NL generation chooses the most compact expression that seems to adequately convey the intended meaning, so it decides on "car man" as the best simple verbalization to match the newly created conceptual blend that it thinks effectively describes the newly created blocks structure. The positive feedback from the user leads to reinforcement of the Atoms and processes that led to the construction of the blocks structure that has been judged beautiful (via importance spreading and SystemActivityTable mining). 50.4 Conclusion The simple situation considered in this chapter is complex enough to involve nearly all the different cognitive processes in the CogPrime system - and many interactions between these processes. This fact illustrates one of the main difficulties of designing, building and testing an artificial mind like CogPrime - until nearly all of the system is build and made to operate in an integrated way, it's hard to do any meaningful test of the system. Testing PLN or MOSES or conceptual blending in isolation may be interesting computer science, but it doesn't tell you much about CogPrime as a design for a thinking machine. According to the CogPrime approach, getting a simple child-like interaction like "build me something with blocks that I haven't seen before" to work properly requires a holistic, integrated cognitive system. Once one has built a system capable of this sort of simple interaction then, according to the theory underlying CogPrime, one is not that far from a system with adult human-level intelligence. And once one has an adult human-level AGI built according to a highly flexible design like CogPrime. given the potential of such systems to self-analyze and self-modify, one is not far off from a dramatically powerful Genius Machine. Of course there will be a lot of work to do to get from a child-level system to an adult-level system - it won't necessarily unfold as "automatically" as seems to happen with a human child, because CogPrime lacks the suite of developmental processes and mechanisms that the young human brain has. But still, a child CogPrime mind capable of doing the things outlined in this chapter will have all the basic components and interactions in place, all the ones that are needed for a much more advanced artificial mind. Of course, one could concoct a narrow-Al system carrying out the specific activities described in this chapter, much more simply than one could build a CogPrime system capable of doing these activities. But that's not the point — the point of this chapter is not to explain how to achieve some particular narrow set of activities "by any means necessary", but rather to explain EFTA00624656

510 50 Build Me Something I Haven't Seen: A CogPrime Thought Experiment how these activities might be achieved within the CogPrime framework, which has been designed with much more generality in mind. It would be worthwhile to elaborate a number of other situations similar to the one described in this chapter, and to work through the various cognitive processes and structures in CogPrime carefully in the context of each of these situations. In fact this sort of exercise has frequently been carried out informally in the context of developing CogPrime. But this book is already long enough, so we will end here, and leave the rest for future works - emphasizing that it is via intimate interplay between concrete considerations like the ones presented in this chapter, and general algorithmic and conceptual considerations as presented in most of the chapters of this book, that we have the greatest hope of creating advanced AGI. The value of this sort of interplay actually follows from the theory of real-world general intelligence presented in Part 1 of the book. Thoroughly general intelligence is only possible given unrealistic computational resources, so real-world general intelligence is about achieving high generality given limited resources relative to the specific classes of environments relevant to a given agent. Specific situations like building surprising things with blocks are particularly important insofar as they embody broader information about the classes of environments relevant to broadly human-like general intelligence. No doubt, once a CogPrime system is completed, the specifics of its handling of the situation described here will differ somewhat from the treatment presented in this chapter. Furthermore, the final CogPrime system may differ algorithmically and structurally in some respects from the specifics given in this book - it would be surprising if the process of building, testing and interacting with CogPrime didn't teach us some new things about various of the topics covered. But our conjecture is that, if sufficient effort is deployed appropriately, then a system much like the CogPrime system described in this book will be able to handle the situation described in this chapter in a roughly similar manner to the one described in this chapter - and that this will serve as a natural precursor to much more dramatic AGI achievements. EFTA00624657

Appendix A Glossary A.1 List of Specialized Acronyms This includes acronyms that are commonly used in discussing CogPrime, OpenCog and related ideas, plus sonic that occur here and there in the text for relatively ephemeral reasons. • AA: Attention Allocation • ADF: Automatically Defined Function (in the context of Genetic Programming) • AF: Attentional Focus • AGI: Artificial General Intelligence • AV: Attention Value • BD: Behavior Description • C-space: Configuration Space • CBV: Coherent Blended Volition • CEV: Coherent Extrapolated Volition • CGGP: Contextually Guided Greedy Parsing • CSDLN: Compositional Spatiotemporal Deep Learning Network • CT: Combo 'Bee • ECAN: Economic Attention Network • ECP: Embodied Communication Prior • EPW : Experiential Possible Worlds (semantics) • FCA: Formal Concept Analysis • FI : Fisher Information • FLM: Frequent Itemset Mining • FOI: First Order Inference • FOPL: First Order Predicate Logic • FOPLN: First Order PLN • FS-MOSES: Feature Selection MOSES (i.e. MOSES with feature selection integrated a la LIFES) • GA: Genetic Algorithms 511 EFTA00624658

512 A Glossary • GB: Global Brain • GEOP: Goal Evaluator Operating Procedure (in a GOLEM context) • GIS: Geospatial Information System • GOLEM: Goal-Oriented LEarning Meta-architecture • GP: Genetic Programming • HOE Higher-Order Inference • HOPLN: Higher-Order PLN • HR: Historical Repository (in a GOLEM context) • HTM: Hierarchical Temporal Memory • IA: (Allen) Interval Algebra (an algebra of temporal intervals) • IRC: Imitation / Reinforcement Correction (Learning) • LIFES: Learning-Integrated Feature Selection • LTI: Long Term Importance • MA: MindAgent • MOSES: Meta-Optimizing Semantic Evolutionary Search • MSH: Mirror System Hypothesis • NARS: Non-Axiomatic Reasoning System • NLGen: A specific software component within OpenCog, which provides one way of dealing with Natural Language Generation • OCP: OpenCogPrime • OP: Operating Program (in a GOLEM context) • PEPL: Probabilistic Evolutionary Procedure Learning (e.g. MOSES) • PLN: Probabilistic Logic Networks • RCC: Region Connection Calculus • RelEx: A specific software component within OpenCog, which provides one way of dealing with natural language Relationship Extraction • SAT: Boolean SATisfaction, as a mathematical / computational problem • SMEPH: Self-Modifying Evolving Probabilistic Hypergraph • SRAM: Simple Realistic Agents Model • STI: Short Term Importance • STY: Simple Truth VAlue • TV: Truth Value • VLTI: Very Long Term Importances • WSPS: Whole-Sentence Purely-Syntactic Parsing A.2 Glossary of Specialized Terms • Abduction: A general form of inference that goes from data describing something to a hypothesis that accounts for the data. Often in an OpenCog context, this refers to the PLN abduction rule, a specific First-Order PLN rule (If A implies C, and B implies C, then maybe A is B), which embodies a simple form of abductive inference. But OpenCog may also carry out abduction, as a general process, in other ways. • Action Selection: The process via which the OpenCog system chooses which Schema to enact, based on its current goals and context. • Active Schema Pool: The set of Schema currently in the midst of Schema Execution. EFTA00624659

A.2 Glossary of Specialized Terms 513 • Adaptive Inference Control: Algorithms or heuristics for guiding PLN inference, that cause inference to be guided differently based on the context in which the inference is taking place, or based on aspects of the inference that are noted as it proceeds. • AGI Preschool: A virtual world or robotic scenario roughly similar to the environment within a typical human preschool, intended for AGIs to learn in via interacting with the environment and with other intelligent agents. • Atom: The basic entity used in OpenCog as an element for building representations. Some Atoms directly represent patterns in the world or mind, others are components of represen- tations. There are two kinds of Atoms: Nodes and Links. • Atom, Frozen: See Atom, Saved • Atom, Realized: An Atom that exists in RAM at a certain point in time. • Atom, Saved: An Atom that has been saved to disk or other similar media, and is not actively being processed. • Atom, Serialized: An Atom that is serialized for transmission from one software process to another, or for saving to disk, etc. • Atom2Link: A part of OpenCogPrime s language generation system, that transforms appropriate Atoms into words connected via link parser link types. • Atomspace: A collection of Atoms, comprising the central part of the memory, of an OpenCog instance. • Attention: The aspect of an intelligent system's dynamics focused on guiding which aspects of an OpenCog system's memory & functionality gets more computational resources at a certain point in time • Attention Allocation: The cognitive process concerned with managing the parameters and relationships guiding what the system pays attention to, at what points in time. This is a term inclusive of Importance Updating and Hebbian Learning. • Attentional Currency: Short Term Importance and Long Term Importance values are implemented in terms of two different types of artificial money, STICurrency and LTICur- rency. Theoretically these may be converted to one another. • Attentional Focus: The Atoms in an OpenCog Atomspace whose ShortTennImportance values lie above a critical threshold (the AttentionalFocus Boundary). The Attention Allo- cation subsystem treats these Atoms differently. Qualitatively, these Atoms constitute the system's main focus of attention during a certain interval of time, i.e. it's a moving bubble of attention. • Attentional Memory: A system's memory of what it's useful to pay attention to, in what contexts. In CogPrime this is managed by the attention allocation subsystem. • Backward Chainer: A piece of software, wrapped in a MindAgent, that carries out back- ward chaining inference using PLN. • CIM-Dynamic: Concretely-Implemented Mind Dynamic, a term for a cognitive process that is implemented explicitly in OpenCog (as opposed to allowed to emerge implicitly from other dynamics). Sometimes a CIM-Dynamic will be implemented via a single MindAgent, sometimes via a set of multiple interrelated MindAgents, occasionally by other means. • Cognition: In an OpenCog context, this is an imprecise term. Sometimes this term means any process closely related to intelligence; but more often it's used specifically to refer to more abstract reasoning/learning/etc, as distinct from lower-level perception and action. • Cognitive Architecture: This refers to the logical division of an AI system like OpenCog into interacting parts and processes representing different conceptual aspects of intelligence. EFTA00624660

514 A Glossary It's different from the software architecture, though of course certain cognitive architectures and certain software architectures fit more naturally together. • Cognitive Cycle: The basic "loop" of operations that an OpenCog system, used to control an agent interacting with a world, goes through rapidly each "subjective moment." Typically a cognitive cycle should be completed in a second or less. It minimally involves perceiving data from the world, storing data in memory, and deciding what if any new actions need to be taken based on the data perceived. It may also involve other processes like deliber- ative thinking or metacognition. Not all OpenCog processing needs to take place within a cognitive cycle. • Cognitive Schematic: An implication of the form "Context AND Procedure IMPLIES goal". Learning and utilization of these is key to CogPrime's cognitive process. • Cognitive Synergy: The phenomenon by which different cognitive processes, controlling a single agent, work together in such a way as to help each other be more intelligent. Typically, if one has cognitive processes that are individually susceptible to combinatorial explosions, cognitive synergy involves coupling them together in such a way that they can help one another overcome each other's internal combinatorial explosions. The CogPrime design is reliant on the hypothesis that its key learning algorithms will display dramatic cognitive synergy when utilized for agent control in appropriate environments. • CogPrime : The name for the AGI design presented in this book, which is designed specifi- cally for implementation within the OpenCog software framework (and this implementation is OpenCogPrime). • CogServer: A piece of software, within OpenCog, that wraps up an Atomspace and a number of MindAgents, along with other mechanisms like a Scheduler for controlling the activity of the MindAgents, and code for important and exporting data from the Atomspace. • Cognitive Equation: The principle, identified in Ben Goertzel's 1994 book "Chaotic Logic", that minds are collections of pattern-recognition elements, that work by iteratively recognizing patterns in each other and then embodying these patterns as new system ele- ments. This is seen as distinguishing mind from "self-organization" in general, as the latter is not so focused on continual pattern recognition. Colloquially this means that "a mind is a system continually creating itself via recognizing patterns in itself." • Combo: The programming language used internally by MOSES to represent the programs it evolves. Schemallodes may refer to Combo programs, whether the latter are learned via MOSES or via some other means. The textual realization of Combo resembles LISP with less syntactic sugar. Internally a Combo program is represented as a program tree. • Composer: In the PLN design, a rule is denoted a composer if it needs premises for generating its consequent. See generator. • CogBuntu: an Ubuntu Linux remix that contains all required packages and tools to test and develop OpenCog. • Concept Creation: A general term for cognitive processes that create new ConceptNodes, PredicateNodes or concept maps representing new concepts. • Conceptual Blending: A process of creating new concepts via judiciously combining pieces of old concepts. This may occur in OpenCog in many ways, among them the explicit use of a ConceptBlending MindAgent, that blends two or more ConceptNodes into a new one. • Confidence: A component of an OpenCog/PLN TruthValue, which is a scaling into the interval 10,11 of the weight of evidence associated with a truth value. In the simplest case (of a probabilistic Simple Truth Value), one uses confidence c = n / (n+k), where n is EFTA00624661

A.2 Clossary of Specialized Terms 515 the weight of evidence and k is a parameter. In the case of an Indefinite Truth Value, the confidence is associated with the width of the probability interval. • Confidence Decay: The process by which the confidence of an Atom decreases over time, as the observations on which the Atom's truth value is bawd become increasingly obsolete. This may be carried out by a special MindAgent. The rate of confidence decay is subtle and contextually determined, and must be estimated via inference rather than simply assumed a priori. • Consciousness: CogPrime is not predicated on any particular conceptual theory of con- sciousness. Informally, the AttentionalFocus is sometimes referred to as the "conscious" mind of a CogPrime system, with the rest of the Atomspace as "unconscious" but this is just an informal usage, not intended to tie the CogPrime design to any particular theory of consciousness. The primary originator of the CogPrime design (Ben Goertzel) tends toward panpsychism, as it happens. • Context: In addition to its general common-sensical meaning, in CogPrime the term Con- text also refers to an Atom that is used as the first argument of a ContextLink. The second argument of the ContextLink then contains Links or Nodes, with TruthValues calculated restricted to the context defined by the first argument. For instance, (ContextLink USA (InheritanceLink person obese )). • Core: The MindOS portion of OpenCog, comprising the Atomspace, the CogServer, and other associated "infrastructural" code. • Corrective Learning: When an agent learns how to do something, by having another agent explicitly guide it in doing the thing. For instance, teaching a dog to sit by pushing its butt to the ground. • CSDLN: (Compositional Spatiotemporal Deep Learning Network): A hierarchical pattern recognition network, in which each layer corresponds to a certain spatiotemporal granularity, the nodes on a given layer correspond to spatiotemporal regions of a given size, and the children of a node correspond to sub-regions of the region the parent corresponds to. Jeff Hawkins's HTM is one example CSDLN, and Itamar Arel's DeSTIN (currently used in OpenCog) is another. • Declarative Knowledge: Semantic knowledge as would be expressed in propositional or predicate logic facts or beliefs. • Deduction: In general, this refers to the derivation of conclusions from premises using logical rules. In PLN in particular, this often refers to the exercise of a specific inference rule, the PLN Deduction rule (A B, B C, therefore A—> C) • Deep Learning: Learning in a network of elements with multiple layers, involving feedfor- ward and feedback dynamics, and adaptation of the links between the elements. An example deep learning algorithm is DeSTIN, which is being integrated with OpenCog for perception processing. • Defrosting: Restoring, into the RAM portion of an Atomspace, an Atom (or set thereof) previously saved to disk. • Demand: In CogPrime's OpenPsi subsystem, this term is used in a manner inherited from the Psi model of motivated action. A Demand in this context is a quantity whose value the system is motivated to adjust. Typically the system wants to keep the Demand between certain minimum and maximum values. An Urge develops when a Demand deviates from its target range. • Deme: In MOSES, an island" of candidate programs, closely clustered together in program space, being evolved in an attempt to optimize a certain fitness function. The idea is that EFTA00624662

516 A Glossary within a dome, programs are generally similar enough that reasonable syntax-semantics correlation obtains. • Derived Hypergraph: The SMEPH hypergraph obtained via modeling a system in terms of a hypergraph representing its internal states and their relationships. For instance, a SMEPH vertex represents a collection of internal states that habitually occur in relation to similar external situations. A SMEPH edge represents a relationship between two SMEPH vertices (e.g. a similarity or inheritance relationship). The terminology "edge /vertex" is used in this context, to distinguish from the 'link / node" terminology used in the context of the Atomspace. • DeSTIN — Deep SpatioTemporal Inference Network: A specific CSDLN created by Itamar Arel, tested on visual perception, and appropriate for integration within CogPrime. • Dialogue: Linguistic interaction between two or more parties. In a CogPrime context, this may be in English or another natural language, or it may be in Lojban or Psynese. • Dialogue Control: The process of determining what to say at each juncture in a dialogue. This is distinguished from the linguistic aspects of dialogue, language comprehension and language generation. Dialogue control applies to Psynese or Lojban, as well as to human natural language. • Dimensional Embedding: The process of embedding entities from some non-dimensional space (e.g. the Atomspace) into an n-dimensional Euclidean space. This can be useful in an Al context because some sorts of queries (e.g. "find everything similar to X", "find a path between X and V) are much faster to carry out among points in a Euclidean space, than among entities in a space with less geometric structure. • Distributed Atomspace: An implementation of an Atomspace that spans multiple com- putational processes; generally this is done to enable spreading an Atomspace across mul- tiple machines. • Dual Network: A network of mental or informational entities with both a hierarchical structure and a heterarchical structure, and an alignment among the two structures so that each one helps with the maintenance of the other. This is hypothesized to be a critical emergent structure, that must emerge in a mind (e.g. in an Atomspace) in order for it to achieve a reasonable level of human-like general intelligence (and possibly to achieve a high level of pragmatic general intelligence in any physical environment). • Efficient Pragmatic General Intelligence: A formal, mathematical definition of general intelligence (extending the pragmatic general intelligence), that ultimately boils down to: the ability to achieve complex goals in complex environments using limited computational resources (where there is a specifically given weighting function determining which goals and environments have highest priority). More specifically, the definition weighted-sums the system's normalized goal-achieving ability over (goal. environment pairs), and where the weights are given by some assumed measure over (goal. environment pairs), and where the normalization is done via dividing by the (space and time) computational resources used for achieving the goal. • Elegant Normal Form (ENF): Used in MOSES, this is a way of putting programs in a normal form while retaining their hierarchical structure. This is critical if one wishes to probabilistically model the structure of a collection of programs, which is a meaningful operation if the collection of programs is operating within a region of program space where syntax-semantics correlation holds to a reasonable degree. The Reduct library, is used to place programs into ENF. EFTA00624663

A.2 Glossary of Specialized Terms 517 • Embodied Communication Prior: The class of prior distributions over (goal, environ- ment pairs), that are imposed by placing an intelligent system in an environment where most of its tasks involve controlling a spatially localized body in a complex world, and in- teracting with other intelligent spatially localized bodies. It is hypothesized that many key aspects of human-like intelligence (e.g. the use of different subsystems for different memory types, and cognitive synergy between the dynamics associated with these subsystems) are consequences of this prior assumption. This is related to the Mind-World Correspondence Principle. • Embodiment: Colloquially, in an OpenCog context, this usually means the use of an AI software system to control a spatially localized body in a complex (usually 3D) world. There are also passible "borderline cases" of embodiment, such as a search agent on the Internet. In a sense any Al is embodied, because it occupies some physical system (e.g. computer hardware) and has some way of interfacing with the outside world. • Emergence: A property or pattern in a system is emergent if it arises via the combination of other system components or aspects, in such a way that its details would be very difficult (not necessarily impossible in principle) to predict from these other system components or aspects. • Emotion: Emotions are system-wide responses to the system's current and predicted state. Dorner's Psi theory of emotion contains explanations of many human emotions in terms of underlying dynamics and motivations, and most of these explanations make sense in a CogPrime context, due to CogPrime's use of OpenPsi (modeled on Psi) for motivation and action selection. • Episodic Knowledge: Knowledge about episodes in an agent's life-history, or the life- history of other agents. CogPrime includes a special dimensional embedding space only for episodic knowledge, easing organization and recall. • Evolutionary Learning: Learning that proceeds via the rough process of iterated differen- tial reproduction based on fitness, incorporating variations of reproduced entities. MOSES is an explicitly evolutionary-learning-based portion of CogPrime; but CogPrime's dynamics as a whole may also be conceived as evolutionary. • Exemplar: (in the context of imitation learning) - When the owner wants to teach an OpenCog controlled agent a behavior by imitation, he/she gives the pet an exemplar. To teach a virtual pet "fetch" for instance, the owner is going to throw a stick, run to it, grab it with his/her mouth and come back to its initial position. • Exemplar: (in the context of MOSES) - Candidate chosen as the core of a new deme, or as the central program within a deme, to be varied by representation building for ongoing exploration of program space. • Explicit Knowledge Representation: Knowledge representation in which individual, easily humanly identifiable pieces of knowledge correspond to individual elements in a knowl- edge store (elements that are explicitly there in the software and accessible via very rapid, deterministic operations) • Extension: In PLN, the extension of a node refers to the instances of the category that the node represents. In contrast is the intension. • Fishgram (Frequent and Interesting Sub-hypergraph Mining): A pattern mining algorithm for identifying frequent and/or interesting sub-hypergraphs in the Atom.space. • First-Order Inference (FOI): The subset of PLN that handles Logical Links not in- volving VariableAtoms or higher-order functions. The other aspect of PLN, Higher-Order Inference, uses Truth Value formulas derived from First-Order Inference. EFTA00624664

518 A Glossary • Forgetting: The process of removing Atoms from the in-RAM portion of AtomSpace, when RAM gets short and they are judged not as valuable to retain in RAM as other Atoms. This is commonly done using the LTI values of the Atoms (removing lowest LTI-Atoms, or more complex strategies involving the LTI of groups of interconnected Atoms). May be done by a dedicated Forgetting MindAgent. VLTI may be used to determine the fate of forgotten Atoms. • Forward Chainer: A control mechanism (MindAgent) for PLN inference, that works by taking existing Atoms and deriving conclusions from them using PLN rules, and then iter- ating this process. The goal is to derive new Atoms that are interesting according to some given criterion. • Frame2Atom: A simple system of hand-coded rules for translating the output of RelEx2Frame (logical representation of semantic relationships using FrameNet relationships) into Atoms. • Freezing: Saving Atoms from the in-RAM AtomSpace to disk. • General Intelligence: Often used in an informal, commonsensical sense, to mean the ability to learn and generalize beyond specific problems or contexts. Has been formalized in various ways as well, including formalizations of the notion of "achieving complex goals in complex environments" and "achieving complex goals in complex environments using limited resources." Usually interpreted as a fuzzy concept, according to which absolutely general intelligence is physically unachievable, and humans have a significant level of general intelligence, but far from the maximally physically achievable degree. • Generalized Hypergraph: A hypergraph with some additional features, such as links that point to links, and nodes that are seen as "containing" whole sub-hypergraphs. This is the most natural and direct way to mathematically/visually model the Atomspace. • Generator: In the Pia design, a rule is denoted a generator if it can produce its consequent without needing premises (e.g. LookupRule, which just looks it up in the AtomSpace). See composer. • Global, Distributed Memory: Memory that stores items as implicit knowledge, with each memory item spread across multiple components, stored as a pattern of organization or activity among them. • Glocal Memory: The storage of items in memory in a way that involves both localized and global, distributed aspects. • Goal: An Atom representing a function that a system (like OpenCog) is supposed to spend a certain non-trivial percentage of its attention optimizing. The goal, informally speaking, is to maximize the Atom's truth value. • Goal, Implicit: A goal that an intelligent system, in practice, strives to achieve; but that is not explicitly represented as a goal in the system's knowledge base. • Goal, Explicit: A goal that an intelligent system explicitly represents in its knowledge has' and expends some resources trying to achieve. Goal Nodes (which may be Nodes or, e.g. hmplicationLLinks) are used for this purpose in OpenCog. • Goal-Driven Learning: Learning that is driven by the cognitive schematic i.e. by the quest of figuring out which procedures can be expected to achieve a certain goal in a certain sort of context. • Grounded Schemallode: See Schemallode, Grounded. • Hebbian Learning: An aspect of Attention Allocation, centered on creating and updating HebbianLinks, which represent the simultaneous importance of the Atoms joined by the HebbianLink. EFTA00624665

A.2 Glossary of Specialized Terms 519 • Hebbian Links: Links recording information about the associative relationship (co- occurrence) between Atoms. These include symmetric and asymmetric HebbianLinks. • Heterarchical Network: A network of linked elements in which the semantic relationships associated with the links are generally symmetrical (e.g. they may be similarity links, or symmetrical associative links). This is one important sort of subnetwork of an intelligent system; see Dual Network. • Hierarchical Network: A network of linked elements in which the semantic relationships associated with the links are generally asymmetrical, and the parent nodes of a node have a more general scope and some measure of control over their children (though there may be important feedback dynamics too). This is one important sort of subnetwork of an intelligent system; see Dual Network. • Higher-Order Inference (HOI): PLN inference involving variables or higher-order func- tion.s. In contrast to First-Order Inference (F00. • Hillclimbing: A general term for greedy, local optimization techniques, including some relatively sophisticated ones that involve "mildly nonlocal" jumps. • Human-Level Intelligence: General intelligence that's "as smart as" human general in- telligence, even if in some respects quite unlike human intelligence. An informal concept, which generally doesn't come up much in CogPrime work, but is used frequently by some other AI theorists. • Human-Like Intelligence: General intelligence with properties and capabilities broadly resembling those of humans, but not necessarily precisely imitating human beings. • Hypergraph: A conventional hypergraph is a collection of nodes and links, where each link may span any number of nodes. OpenCog makes use of generalized hypergraphs (the Atomspace is one of these). • Imitation Learning: Learning via copying what some other agent is observed to do. • Implication: Often refers to an ImplicationLink between two PredicateNodes, indicating an (extensional, intensional or mixed) logical implication. • Implicit Knowledge Representation: Representation of knowledge via having easily humanly identifiable pieces of knowledge correspond to the pattern of organization and/or dynamics of elements, rather than via having individual elements correspond to easily hu- manly identifiable pieces of knowledge. • Importance: A generic term for the Attention Values associated with Atoms. Most com- monly these are STI (short term importance) and LTI (long term importance) values. Other importance values corresponding to various different time scales are also possible. In general an importance value reflects an estimate of the likelihood an Atom will be aseful to the system over some particular future time-horizon. STI is generally relevant to processor time allocation, whereas LTI is generally relevant to memory allocation. • Importance Decay: The process of Atom importance values (e.g. STI and LTI) decreasing over time, if the Atoms are not utilized. Importance decay rates may in general be context- dependent. • Importance Spreading: A synonym for Importance Updating, intended to highlight the similarity with "activation spreading" in neural and semantic networks. • Importance Updating: The CIM-Dynamic that periodically (frequently) updates the STI and LTI values of Atoms based on their recent activity and their relationships. • Imprecise Truth Value: Peter Walley's imprecise truth values are intervals inter- preted as lower and upper bounds of the means of probability distributions in an envelope EFTA00624666

520 A Glossary of distributions. In general, the term may be used to refer to any truth value involving intervals or related constructs, such as indefinite probabilities. • Indefinite Probability: An extension of a standard imprecise probability, comprising a credible interval for the means of probability distributions governed by a given second-order distribution. • Indefinite Truth Value: An OpenCog TruthValue object wrapping up an indefinite prob- ability • Induction: In PLN, a specific inference rule (A —> B, A —> C, therefore B r C). In general, the process of heuristically inferring that what has been seen in multiple examples, will be seen again in new examples. Induction in the broad sense, may be carried out in OpenCog by methods other than PLN induction. When emphasis needs to be laid on the particular PLN inference rule, the phrase "PLN Induction" is used. • Inference: Generally speaking, the process of deriving conclusions from assumptions. In an OpenCog context. this often refers to the PLN inference system. Inference in the broad sense is distinguished from general learning via some specific characteristics, such as the intrinsically incremental nature of inference: it proceeds step by step. • Inference Control: A cognitive process that determines what logical inference rule (e.g. what PLN rule) is applied to what data, at each point in the dynamic operation of an inference process. • Integrative AGI: An AGI architecture, like CogPrime, that relies on a number of different powerful, reasonably general algorithms all cooperating together. This is different from an AGI architecture that is centered on a single algorithm, and also different than an AGI architecture that expects intelligent behavior to emerge from the collective interoperation of a number of simple elements (without any sophisticated algorithms coordinating their overall behavior). • Integrative Cognitive Architecture: A cognitive architecture intended to support inte- grative AGI. • Intelligence: An informal, natural language concept. "General intelligence" is one slightly more precise specification of a related concept; "Universal intelligence" is a fully precise specification of a related concept. Other specifications of related concepts made in the particular context of CogPrime research are the pragmatic general intelligence and the efficient pragmatic general intelligence. • Intension: In PLN, the intention of a node consists of Atoms representing properties of the entity the node represents. • Intentional memory: A system's knowledge of its goals and their subgoaLs. and associa- tions between these goals and procedures and contexts (e.g. cognitive schematics). • Internal Simulation World: A simulation engine used to simulate an external environ- ment (which may be physical or virtual), used by an AGI system as its "mind's eye" in order to experiment with various action' q sequences and envision their consequences, or observe the consequences of various hypothetical situations. Particularly important for dealing with episodic knowledge. • Interval Algebra: Allen Interval Algebra, a mathematical theory of the relationships be- tween time intervals. CogPrime utilizes a fuzzified version of classic Interval Algebra. • IRC Learning (Imitation, Reinforcement, Correction): Learning via interaction with a teacher, involving a combination of imitating the teacher, getting explicit reinforcement signals from the teacher, and having one's incorrect or suboptimal behaviors guided toward betterness by the teacher in real-time. This is a large part of how young humans learn. EFTA00624667

A.2 Glossary of Specialized Terms 521 • Knowledge Base: A shorthand for the totality of knowledge possessed by an intelligent system during a certain interval of time (whether or not this knowledge is explicitly rep- resented). Put differently: this is an intelligence's total memory contents (inclusive of all types of memory) during an interval of time. • Language Comprehension: The process of mapping natural language speech or text into a more "cognitive", largely language-independent representation. In OpenCog this has been done by various pipelines consisting of dedicated natural language processing tools, e.g. a pipeline: text —r Link Parser —> RelEx r RelEx2Frame Frame2Atom Atomspace; and alternatively a pipeline Link Parser —r Link2Atom i Atomspace. It would also be possi- ble to do language comprehension purely via PLN and other generic OpenCog processes, without using specialized language processing tools. • Language Generation: The process of mapping (largely language-independent) cognitive content into speech or text. In OpenCog this has been done by various pipelines consisting of dedicated natural language processing tools, e.g. a pipeline: Atomspace NLGen text; or more recently Atomspace Atom2Link —> surface realization —r text. It would also be possible to do language generation purely via PLN and other generic OpenCog processes, without using specialized language processing tools. • Language Processing: Processing of human language is decomposed, in CogPrime, into Language Comprehension, Language Generation, and Dialogue Control. • Learning: In general, the process of a system adapting based on experience, in a way that increases its intelligence (its ability to achieve its goals). The theory underlying CogPrime doesn't distinguish learning from reasoning, associating, or other aspects of intelligence. • Learning Server: In some OpenCog configurations, this refers to a software server that performs "offline" learning tasks (e.g. using MOSES or hilIclimbing), and is in communica- tion with an Operational Agent Controller software server that performs real-time agent control and dispatches learning tasks to and receives results from the Learning Server. • Linguistic Links: A catch-all tenn for Atoms explicitly representing linguistic content, e.g. WordNode, SentenceNode, CharacterNode. • Link: A type of Atom, representing a relationship among one or more Atoms. Links and Nodes are the two basic kinds of Atoms. • Link Parser: A natural language syntax parser, created by Sleator and Temperley at Carnegie-Mellon University, and currently used as part of OpenCogPrime's natural language comprehension and natural language generation system. • Link2Atom: A system for translating link parser links into Atoms. It attempts to resolve precisely as much ambiguity as needed in order to translate a given assemblage of link parser links into a unique Atom structure. • Lobe: A term sometimes used to refer to a portion of a distributed Atomspace that lives in a single computational process. Often different lobes will live on different machines. • Localized Memory: Memory that stores each item using a small number of closely- connected elements. • Logic: In an OpenCog context, this usually refers to a set of formal rules for translating certain combinations of Atoms into "conclusion" Atoms. The paradigm case at present is the PLN probabilistic logic system, but OpenCog can also be used together with other logics. • Logical Links: Any Atoms whose truth values are primarily determined or adjusted via logical rules, e.g. PLN's hnheritanceLink, SimilarityLink, hnplicationLink, etc. The term isn't usually applied to other links like HebbianLinks whose semantics isn't primarily logic- EFTA00624668

522 A Glossary based, even though these other links can be processed via (e.g. PLN) logical inference via interpreting them logically. • Lojban: A constructed human language, with a completely formalized syntax and a highly formalized semantics, and a small but active community of speakers. In principle this seems an extremely good method for communication between humans and early-stage AGI sys- tems. • Lojban-l-+: A variant of Lojban that incorporates English words, enabling more flexible expression without the need for frequent invention of new Lojban words. • Long Term Importance (LTI): A value associated with each Atom, indicating roughly the expected utility to the system of keeping that Atom in RAM rather than saving it to disk or deleting it. It's possible to have multiple LTI values pertaining to different time scales, but so far practical implementation and most theory has centered on the option of a single LTI value. • LTI: Long Term Importance • Map: A collection of Atoms that are interconnected in such a way that they tend to be commonly active (i.e. to have high STI, e.g. enough to be in the AttentionalFocus, at the same time). • Map Encapsulation: The process of automatically identifying maps in the Atomspace, and creating Atoms that "encapsulate" them; the Atom encapsulation a map would link to all the Atoms in the map. This is a way of making global memory into local memory, thus making the system's memory glocal and explicitly manifesting the "cognitive equation." This may be carried out via a dedicated MapEncapsulation MindAgent. • Map Formation: The process via which maps form in the Atomspace. This need not be explicit; maps may form implicitly via the action of Hebbian Learning. It will commonly occur that Atoms frequently co-occurring in the AttentionalFocus, will come to be joined together in a map. • Memory Types: In CogPrime this generally refers to the different types of memory that are embodied in different data structures or processes in the CogPrime architecture, e.g. declarative (semantic), procedural, attentional, intentional, episodic, sen- sorimotor. • Mind-World Correspondence Principle: The principle that, for a mind to display efficient pragmatic general intelligence relative to a world, it should display many of the same key structural properties as that world. This can be formalized by modeling the world and mind as probabilistic state transition graphs, and saying that the categories implicit in the state transition graphs of the mind and world should be inter-mappable via a high- probability morphism. • Mind OS: A synonym for the OpenCog Core. • MindAgent: An OpenCog software object, residing in the CogServer, that carries out some processes in interaction with the Atomspace. A given conceptual cognitive process (e.g. PLN inference, Attention allocation, etc.) may be carried out by a number of different MindAgents designed to work together. • Mindspace: A model of the set of states of an intelligent system as a geometrical space, imposed by assuming some metric on the set of mind-states. This may be used as a tool for formulating general principles about the dynamics of generally intelligent systems. • Modulators: Parameters in the Psi model of motivated, emotional cognition, that modu- late the way a system perceives, reasons about and interacts with the world. EFTA00624669

A.2 Glossary of Specialized Terms 523 • MOSES (Meta-Optimizing Semantic Evolutionary Search): An algorithm for proce- dure learning, which in the current implementation learns programs in the Combo language. MOSES is an evolutionary learning system, which differs from typical genetic programming systems in multiple aspects including: a subtler framework for managing multiple "demos" or 'islands" of candidate programs: a library, called Reduct for placing programs in Elegant Normal Form; and the use of probabilistic modeling in place of, or in addition to, imitation and crossover as means of determining which new candidate programs to try. • Motoric: Pertaining to the control of physical actuators, e.g. those connected to a robot. May sometimes be used to refer to the control of movements of a virtual character as well. • Moving Bubble of Attention: The Attentional Focus of a CogPrime system. • Natural Language Comprehension: See Language Comprehension • Natural Language Generation: See Language Generation • Natural Language Processing (NLP): See Language Processing • NLGen: Software for carrying out the surface realization phase of natural language gen- eration, via translating collections of RelEx output relationships into English sentences. Was made functional for simple sentences and some complex sentences; not currently under active development, as work has shifted to the related Atom2Link approach to language generation. • Node: A type of Atom. Links and Nodes are the two basic kinds of Atoms. Nodes, math- ematically, can be thought of as "0-ary" links. Some types of Nodes refer to external or mathematical entities (e.g. INordNode, NumberNode); others are purely abstract, e.g. a ConceptNode is characterized purely by the Links relating it to other atoms. Grounded- PredicateNodes and GroundedSchemallodes connect to explicitly represented procedures (sometimes in the Combo language); ungrounded PredicateNodes and Schemallodes are abstract and, like ConceptNodes, purely characterized by their relationships. • Node Probability: Many PLN inference rules rely on probabilities associated with Nodes. Node probabilities are often easiest to interpret in a specific context, e.g. the probability P(cat) makes obvious sense in the context of a typical American house, or in the context of the center of the sun. Without any contextual specification, P(A) is taken to mean the probability that a randomly chosen occasion of the system's experience includes some instance of A. • Novamente Cognition Engine (NCE): A proprietary proto-AGI software system, the predecessor to OpenCog. Many parts of the NCE were open-sourced to form portions of OpenCog, but some NCE code was not included in OpenCog; and now OpenCog includes multiple aspects and plenty of code that was not in NCE. • OpenCog: A software framework intended for development of AGI systems, and also for narrow-AI application using tools that have AGI applications. Co-designed with the Cog- Prime cognitive architecture, but not exclusively bound to it. • OpenCog Prime (OCP): The implementation of the CogPrime cognitive architecture within the OpenCog software framework. • OpenPsi: CogPrime's architecture for motivation-driven action selection, which is based on adapting Dormer's Psi model for use in the OpenCog framework. • Operational Agent Controller (OAC): In some OpenCog configurations, this is a soft- ware server containing a CogServer devoted to real-time control of an agent (e.g. a virtual world agent, or a robot). Background, offline learning tasks may then be dispatched to other software processes, e.g. to a Learning Server. EFTA00624670

524 A Glossary • Pattern: In a CogPrime context, the term "pattern" is generally used to refer to a process that produces some entity, and is judged simpler than that entity. • Pattern Mining: Pattern mining is the process of extracting an (often large) number of patterns from some body of information, subject to some criterion regarding which patterns are of interest. Often (but not exclusively) it refers to algorithms that are rapid or "greedy", finding a large number of simple patterns relatively inexpensively. • Pattern Recognition: The process of identifying and representing a pattern in some substrate (e.g. some collection of Atoms, or some raw perceptual data, etc.). • Patternism: The philosophical principle holding that, from the perspective of engineering intelligent systems, it is sufficient and useful to think about mental processes in terms of (static and dynamical) patterns. • Perception: The process of understanding data from sensors. When natural language is ingested in textual format, this is generally not considered perceptual. Perception may be taken to encompass both pre-processing that prepares sensory data for ingestion into the Atomspace, processing via specialized perception processing systems like DeSTIN that are connected to the Atomspace, and more cognitive-level process within the Atomspace that is oriented toward understanding what has been sensed. • Piagetan Stages: A series of stages of cognitive development hypothesized by develop- mental psychologist Jean Piaget, which are easy to interpret in the context of developing CogPrime systems. The basic stages are: Infantile, Pre-operational, Concrete Operational and Formal. Post-formal stages have been discussed by theorists since Piaget and seem relevant to AGI, especially advanced AGI systems capable of strong self-modification. • PLN: short for Probabilistic Logic Networks • PLN, First-Order: See First-Order hnference • PLN, Higher-Order: See Higher-Order Inference • PLN Rules: A PLN Rule takes as input one or more Atoms (the "premises", usually Links), and output an Atom that is a 'logical conclusion" of those Atoms. The truth value of the consequence is determined by a PLN Formula associated with the Rule. • PLN Formulas: A PLN Formula, corresponding to a PLN Rule, takes the TruthValues corresponding to the premises and produces the TruthValue corresponding to the conclusion. A single Rule may correspond to multiple Formulas, where each Formula deals with a different sort of TruthValue. • Pragmatic General Intelligence: A formalization of the concept of general intelligence, based on the concept that general intelligence is the capability to achieve goals in environ- ments, calculated as a weighted average over some fuzzy set of goals and environments. • Predicate Evaluation: The process of determining the Truth Value of a predicate, embod- ied in a PredicateNode. This may be recursive, as the predicate referenced internally by a Grounded PredicateNode (and represented via a Combo program tree) may itself internally reference other PredicateNodes. • Probabilistic Logic Networks (PLN): A mathematical and conceptual framework for reasoning under uncertainty, integrating aspects of predicate and term logic with extensions of imprecise probability theory. OpenCogPrime's central tool for symbolic reasoning. • Procedural Knowledge: Knowledge regarding which series of actions (or action-combinations) are useful for an agent to undertake in which circumstances. In CogPrime these may be learned in a number of ways, e.g. via PLN or via Hebbian learning of Schema Maps, or via explicit learning of Combo programs via MOSES or hilklimbing. Procedures are represented as Schemallodes or Schema Maps. EFTA00624671

A.2 Glossary of Specialized Terms 525 • Procedure Evaluation/Execution: A general term encompassing both Schema Execu- tion and Predicate Evaluation, both of which are similar computational processes involving manipulation of Combo trees associated with ProcedureNodw. • Procedure Learning: Learning of procedural knowledge, based on any method, e.g. evo- lutionary learning (e.g. MOSES), inference (e.g. PLN), reinforcement learning (e.g. Hebbian learning). • Procedure Node: A Schemallode or PredicateNode • Psi: A model of motivated action and emotion, originated by Dietrich Dorner and further developed by Joscha Bach, who incorporated it in his proto-AGI system MicroPsi. OpenCog- Prime's motivated-action component, OpenPsi, is roughly based on the Psi model. • Psynese: A system enabling different OpenCog instances to communicate without using natural language, via directly exchanging Atom subgraphs, using a special system to map references in the speaker's mind into matching references in the listener's mind. • Psynet Model: An early version of the theory of mind underlying CogPrime, referred to in some early writings on the Webmind AI Engine and Novamente Cognition Engine. The concepts underlying the psynet model are still part of the theory underlying CogPrime, but the name has been deprecated as it never really caught on. • Reasoning: See inference • Reduct: A code library, used within MOSES, applying a collection of hand-coded rewrite rules that transform Combo programs into Elegant Normal Form. • Region Connection Calculus: A mathematical formalism describing a system of basic operations among spatial regions. Used in CogPrime as part of spatial inference to provide relations and rules to be referenced via PLN and potentially other subsystems. • Reinforcement Learning: Learning procedures via experience, in a manner explicitly guided to cause the learning of procedures that will maximize the system's expected future reward. CogPrime does this implicitly whenever it tries to learn procedures that will maxi- mize some Goal whose Truth Value is estimated via an expected reward calculation (where "reward" may mean simply the Truth Value of some Atom defined as "reward"). Goal-driven learning is more general than reinforcement learning as thus defined; and the learning that CogPrime does, which is only partially goal-driven, is yet more general. • RelEx: A software system used in OpenCog as part of natural language comprehension, to map the output of the link parser into more abstract semantic relationships. These more abstract relationships may then he entered directly into the Atomspace, or they may be further abstracted before being entered into the Atomspace, e.g. by RelEx2Frame rules. • RelEx2Frame: A system of rules for translating RelEx output into Atoms, based on the FrameNet ontology. The output of the RelEx2Frame rules make use of the FrameNet library of semantic relationships. The current (2012) RelEx2Frame rule-based is problematic and the RelEx2Frame system is deprecated as a result, in favor of Link2Atom. However, the ideas embodied in these rules may be useful; if cleaned up the rules might profitably be ported into the Atomspace as ImplicationLinks. • Representation Building: A stage within MOSES, wherein a candidate Combo program tree (within a deme) is modified by replacing one or more tree nodes with alternative tree nodes, thus obtaining a new, different candidate program within that deme. This process currently relies on hand-coded knowledge regarding which types of tree nodes a given tree node should be experimentally replaced with (e.g. an AND node might sensibly be replaced with an OR node, but not so sensibly replaced with a node representing a "kick" action). EFTA00624672

525 A Glossary • Request for Services (RFS): In CogPrime's Goal-driven action system, a RFS is a package sent from a Goal Atom to another Atom, offering it a certain amount of STI currency if it is able to deliver the goal what it wants (an increase in its Truth Value). RFS's may be passed on, e.g. from goals to subgoals to sub-subgoals, but eventually an RFS reaches a Grounded Schemallode, and when the corresponding Schema is executed, the payment implicit in the RFS is made. • Robot Preschool: An AGI Preschool in our physical world, intended for robotically em- bodied AGIs. • Robotic Embodiment: Using an AGI to control a robot. The AGI may be running on hardware physically contained in the robot, or may run elsewhere and control the robot via networking methods such as wifi. • Scheduler: Part of the CogServer that controls which processes (e.g. which MindAgents) get processor time, at which point in time. • Schema: A "script" describing a process to be carried out. This may be explicit, as in the case of a GroundedSchemallode, or implicit, as the case in Schema maps or ungrounded Schemallodes. • Schema Encapsulation: The process of automatically recognizing a Schema Map in an Atomspace, and creating a Combo (or other) program embodying the process carried out by this Schema Map, and then storing this program in the Procedure Repository and associating it with a particular Schemallode. This translates distributed, global procedural memory into localized procedural memory. It's a special case of Map Encapsulation. • Schema Execution: The process of "running" a Grounded Schema, similar to running a computer program. Or, phrased alternately: The process of executing the Schema referenced by a Grounded Schemallode. This may be recursive, as the predicate referenced internally by a Grounded Schemallode (and represented via a Combo program tree) may itself internally reference other Grounded Schemallodes. • Schema, Grounded: A Schema that is associated with a specific executable program (either a Combo program or, say, C++ code) • Schema Map: A collection of Atoms, including Schemallodes, that tend to be enacted in a certain order (or set of orders), thus habitually enacting the same process. This is a distributed, globalized way of storing and enacting procedures. • Schema, Ungrounded: A Schema that represents an abstract procedure, not associated with any particular executable program. • Schematic Implication: A general, conceptual name for implications of the form ((Con- text AND Procedure) IMPLIES Goal) • SegSim: A name for the main algorithm underlying the NLGen language generation soft- ware. The algorithm is based on segmenting a collection of Atoms into small parts, and matching each part against memory, to find, for each part, cases where similar Atom- collections already have known linguistic expression. • Self-Modification: A term generally used for Al systems that can purposefully modify their core algorithms and representations. Formally and crisply distinguishing this sort of "strong self-modification" from "mere" learning is a tricky matter. • Sensorimotor: Pertaining to sensory, data, motoric actions, and their combination and intersection. • Sensory: Pertaining to data received by the AGI system from the outside world. In a CogPrime system that perceives language directly as text, the textual input will generally EFTA00624673

A.2 Glossary of Specialized Terms 527 not be considered as "sensory" (on the other hand, speech audio data would be considered as "sensory"). • Short Term Importance: A value associated with each Atom, indicating roughly the expected utility to the system of keeping that Atom in RAM rather than saving it to disk or deleting it. It's possible to have multple LTI values pertaining to different time scales, but so far practical implementation and most theory has centered on the option of a single LTI value. • Similarity: a link type indicating the probabilistic similarity between two different Atoms. Generically this is a combination of Intensional Similarity (similarity of properties) and Extensional Similarity (similarity of members). • Simple Truth Value: a TruthValue involving a pair (s,d) indicating strength (e.g. proba- bility or fuzzy set membership) and confidence d. d may be replaced by other options such as a count n or a weight of evidence w. • Simulation World: See Internal Simulation World • SMEPH (Self-Modifying Evolving Probabilistic Hypergraphs): a style of modeling systems, in which each system is associated with a derived hypergraph • SMEPH Edge: A link in a SMEPH derived hypergraph, indicating an empirically observed relationship (e.g. inheritance or similarity) between two • SMEPH Vertex: A node in a SMEPH derived hypergraph representing a system, indicat- ing a collection of system states empirically observed to arise in conjunction with the same external stimuli • Spatial Inference: PLN reasoning including Atoms that explicitly reference spatial rela- tionships • Spatiotemporal Inference: PLN reasoning including Atoms that explicitly reference spa- tial and temporal relationships • STI: Short hand for Short Term Importance • Strength: The main component of a TruthValue object, lying in the interval I0,1J, refer- ring either to a probability (in cases like InheritanceLink, SimilarityLink, EquivalenceLink, ImplicationLink, etc.) or a fuzzy value (as in MemberLink, EvaluationLink). • Strong Self-Modification: This is generally used as synonymous with Self-Modification, in a CogPrime context. • Subsymbolic: Involving processing of data using elements that have no correspondence to natural language terms, nor abstract concepts; and that are not naturally interpreted as symbolically "standing for" other things. Often used to refer to processes such as perception processing or motor control, which are concerned with entities like pixels or commands like "rotate servomotor 15 by 10 degrees theta and 55 degrees phi." The distinction between "symbolic" and "subsymbolic" is conventional in the history of Al, but seems difficult to formalize rigorously. Logic-based Al systems are typically considered "symbolic", yet • Supercompilation: A technique for program optimization, which globally rewrites a pro- gram into a usually very different looking program that does the same thing. A prototype supercompiler was applied to Combo programs with successful results. • Surface Realization: The process of taking a collection of Atoms and transforming them into a series of words in a (usually natural) language. A stage in the overall process of language generation. • Symbol Grounding: The mapping of a symbolic term into perceptual or motoric entities that help define the meaning of the symbolic term. For instance, the concept "Cat" may be EFTA00624674

528 A Glossary grounded by images of cats, experiences of interactions with cats, imaginations of being a cat, etc. • Symbolic: Pertaining to the formation or manipulation of symbols, i.e. mental entities that are explicitly constructed to represent other entities. Often contrasted with subsymbolic. • Syntax-Semantics Correlation: In the context of MOSES and program learning more broadly, this refers to the property via which distance in syntactic space (distance between the syntactic structure of programs, e.g. if they're represented as program trees) and se- mantic space (distance between the behaviors of programs, e.g. if they're represented as sets of input/output pairs) are reasonably well correlated. This can often happen among sets of programs that are not too widely dispersed in program space. The Reduct library is used to place Combo programs in Elegant Normal Form, which increases the level of syntax-semantics corellation between them. The programs in a single MOSES deme are often closely enough clustered together that they have reasonably high syntax-semantics correlation. • System Activity Table: An OpenCog component that records information regarding what a system did in the past. • Temporal Inference: Reasoning that heavily involves Atoms representing temporal in- formation, e.g. information about the duration of events, or their temporal relationship (before, after, during, beginning, ending). As implemented in CogPrime, makes use of an uncertain version of Allen Interval Algebra. • Truth Value: A package of information associated with an Atom, indicating its degree of truth. Simple'I1uthValue and IndefiniteTruthValue are two common, particular kinds. Multiple truth values associated with the same Atom from different perspectives may be grouped into CompositeTruthValue objects. • Universal Intelligence: A technical term introduced by Shane Legg and Marcus Hutter, describing (roughly speaking) the average capability of a system to carry out computable goals in computable environments, where goal/environment pairs are weighted via the length of the shortest program for computing them. • Urge: In OpenPsi, an Urge develops when a Demand deviates from its target range. • Very Long Term Importance (VLTI): A bit associated with Atoms, which determines whether, when an Atom is forgotten (removed from RAM), it is saved to disk (frozen) or simply deleted. • Virtual AGI Preschool: A virtual world intended for AGI teaching/training/learning, bearing broad resemblance to the preschool environments used for young humans. • Virtual Embodiment: Using as AGI to control an agent living in a virtual world or game world, typically (but not necessarily) a 3D world with broad similarity to the everyday human world. • Webmind AI Engine: A predecessor to the Novamente Cognition Engine and OpenCog, developed 1997-2001 - with many similar concepts (and also some different ones) but quite different algorithms and software architecture EFTA00624675

References 529 References ABS+11. Itamar Arel, $ Berant, T Slonint, A Nfoyal, B Li, and K Chai Sim. Acoustic spatiotemporal modeling using deep machine learning for robust phoneme recognition. In Afeka-AVIOS Speech Processing Conference, 2011. A1183. James F. Allen. Maintaining knowledge about temporal. Intervals CACM, 26:198-3, 1983. AMOI. J. S. Albus and A. M. Nfeystel. Engineering of Mind: An Introduction to the Science of Intelligent Systems. Wiley and Sons, 2001. Ama85. S. Amari. Differential-geometrical methods in statistics. Lecture notes in statistics, 1985. Ama98. S. Amari. Natural gradient works efficiently in learning. Neural Computing, 10:251-276, 1998. ANOO. Shun-ichi Amari and Hiroshi Nagaoka. Methods of information geometry. ANIS, 2000. ARC09a. I. Aral, D. Rose, and R. Coop. Destin: A scalable deep learning architecture with application to high-dimensional robust pattern recognition. Proc. AAA! Workshop on Biologically Inspired Cognitive Architectures, 2009. ARC09b. Itamtu. Arel, Derek Rose, and Robert Coop. A biologically-inspired deep learning architecture with application to high-dimensional pattern recognition. In Biologically Inspired Cognitive Architec- tures, 2009. AAAI Press, 2009. ARKO9. I. Arel, D. Rose, and T. Karnowski. A deep learning architecture comprising homogeneous cortical circuits for scalable spatiotemporal pattern inference. NIPS 2009 Workshop on Deep Learning for Speech Recognition and Related Applications, 2009. Arn69. Rudolf Arnheim. Visual Thinking. University of California Press. Berkeley, 1969. AS94. Rakesh Agrawal and Ra0takrishnan Srilaint. Fast algorithms for mining association rules. In Proc. 20th Int. Conf. Very Large Data Bases, 1994. Ash65. Robert B. Ash. Information Theory. Dover Publications, 1965. Bau04. E. B. Baum. What is Thought? MIT Press, 2004. Bau06. E. Baum. A working hypothesis for general intelligence. In Advances in Artificial General ligence: Concepts, Architectures and Algorithms, 2006. BE07. Neven Boric and Pablo A. Estevez. Genetic programming-based clustering using an information theoretic fitness measure. In Dipti Srinivasan and Lipo Wang, editors, 2007 IEEE Congress on Evolutionary Computation, pages 31-38, Singapore, 25-28 September 2007. IEEE Computational Intelligence Society, IEEE Press. Bel03. Anthony J. Bell. The co-information lattice. Somewhere or other, 2003. Ben94. Brandon Bennett. Spatial reasoning with propositional logics. In Principles of Knowledge Repre- sentation and Reasoning: Proceedings of the 4th International Conference (KR94), pages 51-62. Morgan Kaufmann, 1994. BF97. A Blum and M Furst. Fast planning through planning graph analysis. Artificial intelligence, 1997. BH10. Bundzel and Hashimoto. Object identification in dynamic images based on the memory-prediction theory of brain function. Journal of Intelligent Learning Systems and Applications, 2-4, 2010. Bic08. Derek Bickerton. Bastard Tongues. Hill and Wang, 2008. BKLO6. Aline Beygelzimer, Sham Kakade, and John Langford. Cover trees for nearest neighbor. In Proc. International Conference on Machine Learning, 2006. BL99. Avrim Blum and John Langford. Probabilistic planning in the graphplan framework. In 5th European Conference on Planning (ECP '99), 1999. Bor05. Christian Borgelt. Keeping things simple: Finding frequent item sets by recursive elimination. In Workshop on Open Source Data Mining Software (OSDM'05). Chicago IL, pages 66-70. 2005. Car06. Pereira Francisco Cara. Creativity and Artificial Intelligence: A Conceptual Blending Approach, Applications of Cognitive Linguistics. Amsterdam: Mouton de Gruyter, 2006. Cas04. N. L. Cassimatis. Grammatical processing using the mechanisms of physical inferences. In Pro- ceedings of the Twentieth-Sixth Annual Conference of the Cognitive Science Society. 2004. CB00. W. H. Calvin and D. Bickerton. Lingua er Machin. MIT Press, 2000. CFH97. Eliseo Clementini, Paolino Di Felice, and Daniel HernAindez. Qualitative representation of posi- tional information. Artificial Intelligence, 95:317-356, 1997. COPH09. Lucio Coelho, Ben Goertzel, Cassio Pennachin, and Chris Howard. Classifier ensemble based analysis of a genome-wide snp dataset concerning late-onset alzheimer disease. In Proceedings of 8th IEEE International Conference on Cognitive Informatics., 2009. Cha0s. Gregory Chaitin. Algorithmic Information Theory. Cambridge University Press, 2008. EFTA00624676

530 A Glossary Cha09. Mark Changizi. The Vision Revolution. l3enBella Books, 2009. Che97. K. Chellapilla. Evolving computer programs without subtree crossover. IEEE Transactions on Evolutionary Computation, 1997. Coh95. A.C. Cohn. A hierarchical representation of qualitative shape based on connection and convexity. In Proc COSIT95, LNCS, pages 311-326. Springer Verlag, 1995. Cox61. Richard Cox. The Algebra of Probable Inference. Johns Hopkins University Press, 1961. CS10. Shay B. Cohen and Noah A. Smith. Covariance in unsupervised learning of probabilistic grammars. Journal of Machine Learning Research, 11:3117-3151, 2010. CSZ06. Olivier Chapelk, Bernhard Schakopf, and Alexander Zien. Semi-Supervised Learning. MIT Press, 2006. CXYNI05. Yun Chi, Yi Xia, Yirong Yang, and Richard Ft Muntz. Mining closed and maximal frequent subtrees from databases of labeled rooted trees IEEE Trans. Knowledge and Data Engineering. 2005. Dab99. A.C. Dabak. A Geometry for Detection Theory. PhD Thesis, Rice U., 1999. Dea98. Terrence Deacon. The Symbolic Species. Norton, 1998. dF37. Bnmo de Finetti. La prevision: ses lois logiques, sea sources subjectives,. Annetta de l'In.stitut Henri Poincard, 1937. DP09. Yassine Djouadi and Henri Prade. Interval-valued fuzzy formal concept analysis. In ISMS '09: Proc. of the 18th International Symposium on Foundations of Intelligent Systems, pages 592-601, Berlin, Heidelberg, 2009. Springer-Verlag. dS77. Ferdinand de Saussure. Course in General Linguistics. Flmtana/Collins, 1977. Orig. published 1916 as "emirs de linguistique generale". EBJ+97. J. Elman, E. Bates, M. Johnson, A. Karmiloff-Smith, D. Parisi. and K. Plunkett. Rethinking Innateness: A Connectionist Perspective on Development. MIT Press, 1997. Ede93. Gerald Edelman. Neural darwinism: Selection and reentrant signaling in higher brain function. Neuron, 10, 1993. PF92. Christian Freksa and Robert Milton. Temporal reasoning based on semi-intervals. Artificial Intel- ligence, 54(1-2):199 - 227, 1992. PLI2. Jeremy Fishel and Gerald Loeb. Bayesian exploration for intelligent identification of textures. Frontiers in Neurorobotics 6-4, 2012. Pri98. Roy. FYieden. Physics from Fisher Information. Cambridge U. Press, 1998. PT02. G. Pauconnier and M. Turner. The Way We Think: Conceptual Blending and the Mind's Hidden Complexities. Basic, 2002. Gar00. Peter Gardenfors. Conceptual spaces: the geometry of thought. MIT Press, 2000. GBK04. S. Gustafson. E. K. Burke, and G. Kendall. Sampling of unique structures and behaviours in genetic programming. In European Conf. on Genetic Programming. 2004. CCPM06. Ben Goertzel, Lucio Coelho, Cassio Pennachin, and Mauricio Mudada. Identifying Complex Bio- logkal Interactions based on Categorical Gene Expression Data. In Proceedings of Conference on Evolutionary Computing. Vancouver CA. 2006. CE01. Roop Goyal and Max Egenhofer. S" 'larity in cardinal directions. In in Proc. of the Seventh International Symposium on Spatial and Temporal Databases, pages 36-55. Springer-Verlag, 2001. Cea05. Ben Coertzel and et al. Combinations of single nucleotide polymorphisms in neuroendocrine effector and receptor genes predict chronic fatigue syndrome. Pharmacogenatnics, 2005. GEA08. Ben Gone! and Cassie. Pennachin Et Al. An integrative methodology, for teaching embodied non-linguistic agents, applied to virtual animals in second life. In Proc.of the First Conf. on AGL IOS Press, 2008. Ceal3. Ben Goertzel and et al. The cogprime architecture for embodied artificial general intelligence. In Proceedings of IEEE Symposium on Human-Level Al, Singapore, 2013. CCC+ 11. Ben Goertzel, Nil Geisweiller, Lucio Coelho, Predrag Janicic, and Cassio Pennachin. Real World Reasoning. Atlantis, 2011. CH11. N Garg and J Henderson. Temporal restricted boltzmann machines for dependency parsing. In Proc. ACL, 2011. GIII. B. Coertzel and M. Ikle. Steps toward a geometry of mind. In J Schmidhuber and K Thorisson, editors, Subm.to ACI-11. Springer, 2011. CICH08. B. Goertzel, M. Ikle, I. Goertzel, and A. Heljakka. Probabilistic Logic Networks. Springer, 2008. CKD89. D. E. Goldberg, B. Korb, and K. Deb. Messy genetic algorithms: Motivation, analysis, and first results. Complex Systems, 1989. EFTA00624677

References 531 CLI0. Ben Coertzel and Ruiting Lian. A probabilistic characterization of fuzzy semantics. Proc. of ICAI-10, Beijing, 2010. CLdG+ 10. Ben Coertzel, Ruiting Lien, Hugo de Gar's, Shuo Chen, and hamar Arel. World survey of artificial brains, part ii: Biologically inspired cognitive architectures. Neurocomputing, April 2010. GM11108. B. Goertzel, I. Coertzel M. lkle, and A. Heljakka. Probabilistic Logic Networks. Springer, 2008. GN02. Alfonso Gerevini and Bernhard Nebel. Qualitative spatio-temporal reasoning with rcc-8 and alien's interval calculus: Computational complexity. In Frank van Harmelen, editor, ECAI, pages 312-316. IOS Press, 2002. Goe94. Ben Coertzel. Chaotic Logic. Plenum, 1994. Coe06. Ben Coertzel. The Hidden Pattern. Brown Walker, 2006. Coe08a. B. Coertzel. The pleasure algorithm. groups.google.com/group/opencog/files, 2008. Coe08b. Ben Coertzel. A pragmatic path toward endowing virtually-embodied ais with human-level lin- guistic capability. IEEE World Congress on Computational Intelligence (Wee!), 2008. Goel0a. Ben Goertzel. Infinite-order probabilities and their application to modeling self-referential seman- tics. In Proceedings of Conference on Advanced Intelligence 2010, Beijing, 2010. Goel0b. Ben et al Coertzel. A general intelligence oriented architecture for embodied natural language processing. In Proc. of the Third Conf. on Artificial General Intelligence (ACI-10). Atlantis Press, 2010. Cool 1 a. B Coertzel. Integrating a compositional spatiotemporal deep learning network with symbolic representation/reasoning within an integrative cognitive architecture via an intermediary semantic network. In Proceedings of AAA! Symposium on Cognitive Systems„ 2011. Goel 1 b. Ben Goertzel. Imprecise probability as a linking mechanism between deep learning, symbolic cognition and local feature detection in vision processing. In Proc. of AC!-11, 2011. CPPG116. Ben Coertzel, Hugo Pinto, Cassio Pennachin, and Izabela Freire Coertzel. Using dependency pars- ing and probabilistic inference to extract relationships between genes, proteins and malignancies implicit among multiple biomedical research abstracts. In Proc. of Bio-NLP 2006, 2006. CR00. Alfonso Gerevini and Jochett Renz. Combining topological and size information for spatial reason- ing. Artificial Intelligence, 137:2002, 2000. GSW05. Bernhard Canter, Gerd Stumme, and Rudolf Wille. Formal Concept Analysis: Foundations and Applications. Springer-Verlag, 2005. HB06. Jeff Hawkins and Sandra Blakeslee. On Intelligence. Brown Walker, 2006. HDY+12. Geoffrey Hinton, Li Deng, bong Yu, George bald, Abdel rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, and Tara Sainathand Brian Kingsbury. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine, 2012. 141407. Barbara Hammer and Pascal Ritzier, editors. Perspectives of Neural-Symbolic Integration. Studies in Computational Intelligence, Vol. 77. Springer, 2007. Hi189. Daniel Hillis. The Connection Machine. MIT Press, 1989. HK02. David Harel and Yehuda Koren. Graph Drawing by High-Dimensional Embedding. 2002. Hob78. J. Hobbs. Resolving pronoun references. Lingua, 44:311-338, 1978. Hof79. Douglas Holstadter. Codel, Escher, Bach: An Eternal Golden Braid. Basic, 1979. Hol75. J. R. Holland. Adaptation in Natural and Artificial Systems. University of Michigan Press, 1975. Hud84. Richard Hudson. Word Grammar. Oxford: Blackwell, 1984. Htic190. Richard Hudson. English Word Grammar. Blackwell Press, 1990. Hud07a. Richard Hudson. Language Networks. The new Word Grammar. Oxford University Press, 2007. Hud07b. Richard Hudson. Language Networks: The New Word Grammar. Oxford Linguistics, 2007. Hut99. G. Hutton. A tutorial on the universality and expressiveness of fold. Journal of anctional Programming, 1999. Hut05a. Marcus Hutter. Universal Artificial Intelligence: Sequential Decisions based on Algorithmic Prob- ability. Springer, 2005. Hut05b. Marcus Hutter. Universal Artificial Intelligence: Sequential Decisions based on Algorithmic Prob- ability. Springer, 2005. HWP03. Jun Huan, Wei Wang, and Jan Prins. Efficient mining of frequent subgraph in the presence of isomorphism. In Proceedings of the 3rd IEEE International Conference on Data Mining (ICDM), pages 549-552. 2003. Jac03. Ray Jackendoff. Foundations of Language: Brain, Meaning, Grammar, Evolution. Oxford Uni- versity Press, 2003. EFTA00624678

532 A Glossary JLO8. D. J. Jilk and C. Lebiere. and o'reilly. R. C. and Anderson, J. R. (2008). SAL: An explicitly pluralistic cognitive architecture. Journal of Experimental and Theoretical Artificial Intelligence, 20:197-218, 2008. Joh05. Mark Johnson. LATEX: A Developmental Cognitive Neuroscience. Wiley-Blackwell, 2005. Jolla. I. T. Joliffe. Principal Component Analysis. Springer, 2010. KA95. J. R. Koza and D. Andre. Parallel genetic programming on a network of transputers. Technical report, Stanford University, 1995. KARIO. Tom Karnowski, hamar Arel, and D. Rose. Deep spatiotemporal feature learning with application to image classification. In The 9th International Conference on Machine Learning and Applications (ICAI LA '10), 2010. KK01. Michihiro Kuramochi and George Karypis. Ftequent subgraph discovery. In Proceedings of the 2001 IEEE International Conference on Data Mining, pages 313-320. 2001. KM04. Dan Klein and Christopher D. Manning. Corpus-based induction of syntactic structure: Models of dependency and constituency. In ACL '04 Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, pages 479-486. Association for Computational Linguistics, 2004. Koh01. Teuvo Kohonen. Self-Organizing Maps. Springer, 2001. Koz92. J. R. Koza. Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, 1992. Koz94. J. R. Koza. Genetic Programming II: Automatic Discovery of Reusable Programs. MIT Press, 1994. KSPC13. Dimitri Kartsaklis, Mehrnoosh Sadrzadeh, Stephen Putman, and Bob Coecke. Reasoning about meaning in natural language with compact closed categories and frobenius algebras. 2013. Kur09. Yohei Kuria. 9-intersection calculi for spatial reasoning on the topological relations between multi-domain objects. USA, June 2009. Kur12. Ray Kurzweil. How to Create a Mind. Viking, 2012. LA93. C Lebiere and J R Anderson. A connectionist implementation of the act-r production system. In Proceedings of the Fifteenth Annual Conference of the Cognitive Science Society, 1993. Lai12. John E Laird. The Soar Cognitive Architecture. MIT Press, 2012. LBHIO. Jens Lehmann, Sebastian Bader, and Pascal Hitzler. Extracting reduced logic programs from artificial neural networks. Applied Intelligence, 2010. LDA05. Ramon Lopez, Cozar Delgado, and Masahiro Araki. Spoken, Multilingual and Afziltimodal Dialogue Systems: Development and Assessment. Wiley, 2005. Lamb°. B. Lemoine. Nlgen2: a linguistically plausible, general purpose natural language generation system. http://www.louisiana.edu/??ba12277/WLGen2, 2010. Lev94. L. Levin. Randomness and ttondeterminism. In The International Congress of Mathematicians, 1994. LGEIO. Ruiting Lien, Ben Goertzel, and Al Et. Language generation via glocal similarity matching. Neurocomputing, 2010. LOKI' 12. Ruiting Lien, Ben Goertzel, Shujing Ke, Jade OONeill, Keyvan Sadeghi, Simon Shin, Dingjie Wang, Oliver Watkins. and Gino Yu. Syntax-semantic mapping for general intelligence: Language comprehension as hypergraph homomorphism, language generation as constraint satisfaction. In Artificial General Intelligence: Lecture Notes in Computer Science Volume 7716. Springer, 2012. LKP+05. Sung Hee Lee, Junggon Kim, Frank Chongwoo Park, Munsang Kim, and James E. Bobrow. Newton-type algorithms for dynamics-based robot movement optimization. IEEE Transactions on Robotics, 21(4):657-667, 2005. LLR09. Weiming Liu. Sanjiang Li, and Jochen Rens. Combining rcc-8 with qualitative direction calculi: Algorithms and complexity. In IJCAI, 2009. LXIDK07. Thomas K. Landauer, Danielle S. McNamara, Simon Dennis, and Walter Kintsch. Handbook of Latent Semantic Analysis. Psychology Press, 2007. LNOO. George Lakoff and Rafael Nunez. Where Mathematics Comes From. Basic Books, 2000. Loo06. Moshe Looks. Competent Program Evolution. PhD Thesis, Computer Science Department, Wash- ington University, 2006. Loo07a. M. Looks. On the behavioral diversity of random programs. In Genetic and evolutionary compu- tation conference, 2007. Loo07b. M. Looks. Scalable estimation-of-distribution program evolution. In Genetic and evolutionary computation conference, 2007. EFTA00624679

References 533 Loo07c. Moshe Looks. Meta-optimizing semantic evolutionary search. In Hod Lipson, editor, Genetic and Evolutionary Computation Conference, CECCO 2007, Proceedings, London, England, UK, July 7-11, 2007, page 626. ACM, 2007. Low99. David Lowe. Object recognition front local scale-invariant features. In P MC. of the International Conf. on Computer Vision, pages 1150-1157, 1999. LP0I. Dekang Lin and Patrick Pantel. Dirt: Discovery of inference rules from text. In Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'01), pages 323428. ACM Press, 2001. LP02. W. B. Langdon and IL Poli. Foundations of Genetic Programming. Springer-Verlag, 2002. Mai00. Monika Maidl. The common fragment of ctl and lel. In In IEEE Symposium on Foundations of Computer Science, pages 643-652, 2000. May04. M. T. Maybury. New Directions in Question Answering. MIT Press, 2004. Mea07. Reiman E M and et al. Gab2 alleles modify alzheimer's risk in apoe e4 carriers. Neuron 54 (5), 2007. Mih05. Rada Mihalcea. Unsupervised large-vocabulary word sense disambiguation with graph-based algo- rithms for sequence data labeling. In HLT Proceedings of the conference on Human Language Technology end Empirical Alethods in Natural Language Processing, pages 411-418, Morristown, NJ, USA, 2005. Association for Computational Linguistics. Mih07. Rada Nlihalcea. Wand sense disambiguation. Encyclopedia of Machine Learning. Springer-Verlag, 2007. Min88. Marvin Minsky. The Society of Mind. MIT Press, 1988. Mit96. Steven !Althorn The Prehistory of Mind. Thames and Hudson, 1996. MS99. Christopher Manning and Heinrich Scheutze. Foundations of Statistical Natural Language Pro- cessing. MIT Press, 1999. NITF04. Rada Nlihalcea, Paul Tarau, and Elizabeth Figa. Pagerank on semantic networks, with application to word sense disambiguation. In COLIN° '04: Proceedings of the 20th international confer- ence on Computational Linguistics, Morristown, NJ, USA, 2004. Association for Computational Linguistics. OCC90. Andrew Ortony, Gerald Clore, and Allan Collins. The Cognitive Structure of Emotions. Cambridge University Press, 1990. O1s95. J. R. Olsson. Inductive functional programming using incremental program transformation. Arti- ficial Intelligence, 1995. PAF00. H. Park, S. Antari, and K. Fukumizu. Adaptive natural gradient learning algorithms for various stochastic models. Neural Computing, 13:755-764, 2000. Pal04. Girish Keshav Palshficar. Fuzzy region connection calculus in finite discrete space domains. Appl. Soft Comput., 4(1):13-23, 2004. PCP00. Papageorgiou, C., and T. Poggio. A trainable system for object detection, volume 38-1. Intl. .) ComputerVision, 2000. PD09. Hoifung Poon and Pedro Domingos. Unsupervised semantic parsing. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1-10, Singapore, August 2009. Association for Computational Linguistics. Pei34. C. Peirce. Collected papers: Volume V. Pragmatism end pragmaticism. Harvard University Press. Cambridge MA., 1934. Pel05. Martin Pelikan. Hierarchical Bayesian Optimization Algorithm: Toward a New Generation of Evolutionary Algorithms. Springer, 2005. PJ88a. Pearl and J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufman. 1988. PJ88b. Steven Pinker and .1acquesMehler. Connections and Symbols. MIT Press, 1988. Pro13. The Univalent Foundations Program. Homology Type Theory: Univalent Foundations of Mathe- matics. Institute for Advanced Study, 2013. RCC93. D. A. Randell, Z. Cui, and A. G. Cohn. A spatial logic based on regions and connection. 1993. Rcs99. J. Reece. Genetic programming acquires solutions by combining top-down and bottom-up refine- ment. In Foundations of Generic Programming. 1999. Row90. John Rowan. Subpersonalities: The People Inside Us. Routledge Press, 1990. RVG05. Mike Ross, Linos Vepstas, and Ben Goertzel. Rolex semantic relationship extractor. http://opencog.org/wilci/RelEx, 2005. EFTA00624680

534 A Glossary Sch06. .1. Schmidhuber. Code! machines: Fully Self-referential Optimal Universal Self-improvers. In B. Go- ertzel and C. Pennachin, editors, Artificial General Intelligence, pages 119-226. 2006. SDCCK08a. Steven Schockaert, Martine De Cock. Chris Cornelis, and Etienne E. Kerne. Fuzzy region con- nection calculus: An interpretation based on closeness. Int. J. Approx. Reasoning, 48(1):332-347, 2008. SDCCK08b. Steven Schockaert, Martine De Cock, Chris Cornelis, and Etienne E. Kerne. Fuzzy region connec- tion calculus: Representing vague topological information. Int. J. Approx. Reasoning, 48(1):314- 331, 2008. SM07. Ravi Sinha and Rada Mihalcea. Unsupervised graph-basedword sense disambiguation using mea- sures of word semantic similarity. In ICSC '07: Proceedings of the International Conference on Semantic Computing, pages 363-369, Washington, DC, USA, 2007. IEEE Computer Society. SM09. Ravi Sinha and Rada Mihalcea. Unsupervised graph-based word sense disambiguation. In Nicolas Nicolov and John Benjamins Ruslan Xlitkov, editors, Current Issues in Linguistic Theory: Recent Advances in Natural Language Processing. 2009. SMI97. F-R. Sinot, Fernandez M., and Mackie I. Efficient reductions with director strings. Evolutionary Computation, 1997. SMK12. Jeremy Stober, Risto Miikkulainen, and Benjamin Kuipers. Learning geometry from sensorimo- tor experience. In Proceedings of the First Joint Conference on Development and Learning and Epigenetic Robotics, 2012. Sol64a. Ray Solomonoff. A Formal Theory of Inductive Inference, Part I. Information and Control, 1964. Sol64b. Ray Solomonoff. A Formal Theory of Inductive Inference, Part IL. Information and Control, 1964. Spe96. L. Spector. Simultaneous evolution of programs and their control structures. In Advances in Genetic Programming 2. MIT Press, 1996. SR0t. Murray Shanahan and David A Randell. A logic-based formulation of active visual perception. In Knowledge Representation, 2004. 5S03. R. P. Salustowicz and J. Schmidhuber. Probabilistic incremental program evolution. Lecture Notes in Computer Science vol. 2706, 2003. ST91. Daniel Sleator and Davy Tetnperley. Parsing english with a link grammar. Technical report, Carnegie Mellon University Computer Science technical report CMU-CS-91-196, 1991. ST93. Daniel Sleator and Davy Temperley. Parsing english with a link grammar. Third International Workshop on Parsing Technologies., 1993. SV99. A. J. Storkey and R. Valabregue. The basins of attraction of a new hopfield learning rule. Neural Networks, 12:869-876, 1999. SW05. Reza Shadmehr and Steven P. Wise. The Computational Neurobiology of Reaching and Pointing : A Foundation for Motor Learning. MIT Press, 2005. SWM90. Timothy Starkweather, Darrell Whitley, and Keith Mathias. Optimization using disributed genetic algorithms. In Parallel Problem Solving from Nature, Edited by H Schwefel and R Manner, 1990. 5204. R. Sun and X. Zhang. Top-down versus bottom-up learning in cognitive skill acquisition. Cognitive Systems Research, 5, 2004. Tes59. Lucien TesniEre. Elements de syntaxe structurale. Klincksieck, Paris, 1959. Tom03. Michael Tomasello. Constructing a Language: A Usage-Based Theory of Language Acquisition. 2003. TSH11. Mohamad Tarifi, Moore Sitharam, and Jeffery Ho. Learning hierarchical sparse representations using iterative dictionary learning and dimension reduction. In Proc. of RICA 2011, 2011. TVCC05. M. Tomassini, L. Vanneschi, P. Collard, and M. Clergue. A study of fitness distance correlation as a difficulty measure in genetic programming. Evolutionary Computation, 2005. VK94. T. Veale and XI. T. Keane. Metaphor and Memory and Meaning in Sapper: A Hybrid Model of Metaphor Interpretation. Proceedings of the workshop on Hybrid Connectionist Systems of ECAI94, at the llth European Conference on Artificial Intelligence, 1994. VO07. Tony Veak and Diannuid O'Donoghue. Computation and Blending. Cognitive Linguistics, 2007. Wah06. Wolfgang Wahlster. SmartKom: Foundations of Multimodel Dialogue Systems. Springer, 2006. Wan06. Pei Wang. Rigid Flexibility: The Logic of Intelligence. Springer, 2006. WF05. Ian Witten and Eibe Frank. Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann, 2005. Win95. Stephan Winter. Topological relations between discrete regions. In Advances in Spatial Databasesd 4th International Symposium, SSD -5, pages 310-327. Springer, 1995. EFTA00624681

References 535 Win00. Stephan Winter. Uncertain topological relations between imprecise regions. Journal of Geograph- ical Information Science, 14(5):411-430, 2000. WKB05. Nice Van De Weghe, Bart Kuijpers, and Peter Bogaert. A qualitative trajectory calculus and the composition of its relations. In Proc. of GeoS, pages 60-76. Springer-Verlag, 2005. Yan10. King-Yin Yen. A fuzzy-probabilistic calculus for vagueness. In Unpublished manuscript, 2010. Subm. YKL+04. Sanghoon We, Jinwook Kim, Sung Hee Lee, Frank Chongwoo Park, Wooram Park, Junggon Kim, Changbeom Park, mid Intaeck Yeo. A modular object-oriented framework for hierarchical multi- resolution robot simulation. Robotics, 22(2):141-154, 2004. Yur98. Denis Yuret. Discovery of Linguistic Relations Using Lexical Attraction. PhD thesis, MIT, 1998. ZHIO. Ruoyu Zou and Lawrence B. Holder. Frequent subgraph mining on a single large graph using sampling techniques. In International Conference on Knowledge Discovery and Data Mining archive. Proceedings of the Eighth Workshop on Mining and Learning with Graphs. Washington DC, pages 171-178, 2010. ZLLY08. Xiaotong Zhang, Weiming Liu, Sanjiang Li, and Xlingsheng Ying. Reasoning with cardinal di- rections: an efficient algorithm. In AAAI'08: Proc. of the 28rd national conference on Artificial intelligence, pages 387-392. AAAI Press, 2008. ZMO6. Song-Chun Zhu and David Mumford. A stochastic grammar of images. Foundations and Trends in Computer Graphics and Vision, 2006. EFTA00624682