Representational Limits in Cognitive Architectures Antonio Lieto University of Turin, Department of Computer Science, Italy ICAR-CNR, Palermo, Italy lieto.antonio@gmail.com http://www.antoniolieto.net Abstract—This paper proposes a focused analysis on some general content. This means that the knowledge embedded and problematic aspects concerning the knowledge level in General processed in such architectures is usually very limited, ad-hoc Cognitive Architectures (CAs). In particular, it addresses the built, domain specific, or based on the specific tasks they have problems regarding both the limited size and the homogeneous to deal with. Thus, every evaluation of the artificial systems typology of the encoded (and processed) conceptual knowledge. relying upon them, is necessarily task-specific and do not As a possible way out to face, jointly, these problems, this involve not even the minimum part of the full spectrum of contribution discusses the possibility of integrating external, but processes involved in the human cognition when the architecturally compliant, cognitive systems into the knowledge “knowledge” comes to play a role. As a consequence, the representation and processing mechanisms of the CAs. structural mechanisms that the CAs implement concerning knowledge processing tasks (e.g. that ones of retrieval, Keywords—cognitive architectures, knowledge representation, knowledge level, common-sense reasoning. learning, reasoning etc.) can be only loosely evaluated, and compared w.r.t. that ones used by humans in similar I. INTRODUCTION knowledge-intensive situations. In other words: from an epistemological perspective, the explanatory power of their The research on Cognitive Architectures (CAs) is a wide computational simulation is strongly affected [7,8]. Such and active area involving a plethora of disciplines such as knowledge limitation, in our opinion, does not allow to obtain Cognitive Science, Artificial Intelligence, Robotics and, more significant advancements in the cognitive science research recently, the area of Computational Neuroscience. CAs have about how the humans heuristically select and deal with the been historically introduced i) “to capture, at the computational huge amount of knowledge that possess when they have to level, the invariant mechanisms of human cognition, including make decisions, reason about a given situation or, more in those underlying the functions of control, learning, memory, general, solve a particular cognitive task involving several adaptivity, perception and action” [1] and ii) to reach human dimensions of analysis. This problem, as a consequence, also level intelligence, also called General Artificial Intelligence, by limits the advancement of the research in the area of General means of the realization of artificial artifacts built upon them. Artificial Intelligence of cognitive inspiration. During the last decades many cognitive architectures have been realized, - such as ACT-R [2], SOAR [3] etc. - and have been The “content” limit of the cognitive architectures has been widely tested in several cognitive tasks involving learning, recently pointed out in literature [1] and some technical reasoning, selective attention, multimodal perception, solutions for filling this “knowledge gap” have been proposed recognition etc. Despite the recent developments, however, in [9]. In particular the use of ontologies and of semantic the last decades the importance of the “knowledge level” [4] formalisms and resources (such as DBPedia) has been seen as a has been historically and systematically downsized by this possible solution for providing effective content to the research area, whose interests have been mainly based on the structural knowledge modules of the cognitive architectures. analysis and the development of mechanisms and the processes Some initial efforts have been done in this sense but cover only governing human and (artificial) cognition. The knowledge part of the “knowledge problem” in CAs (i.e. the one level in CAs, however, presents several problems that may concerning the limited “size” of the adopted knowledge bases). affect the overall heuristic and epistemological value of such However, also these solutions, do not address another relevant artificial general systems and therefore deserves more aspect affecting the knowledge level of CAs: namely, the attention. problem concerning the “knowledge homogeneity” issue. In other terms: the type of knowledge represented and II. TWO PROBLEMS FOR THE KNOWLEDGE LEVEL IN CAS manipulated by most CAs (including those provided with Handling a huge amount of knowledge, and selectively extended knowledge modules) is usually homogeneous in retrieve it according to the needs emerging in different nature. It mainly covers, in fact, only the so called “classical” situational scenarios, represents an important aspect of human part of conceptual information (that one representing concepts intelligence. For this task humans adopt a wide range of in terms of necessary and sufficient information and compliant heuristics [5] due to their “bounded rationality” [6]. Currently, with ontological semantics (see [10]) on these aspects). On the however, the Cognitive Architectures are not able, de facto, to other hand, the so called “common-sense” conceptual deal with complex knowledge structures that can be even components of our knowledge (i.e. those that, based on the slightly comparable to the knowledge heuristically managed by results from the cognitive science, allow to characterize humans. In other terms: CAs are general structures without a concepts in terms of “prototypes”, “exemplars” or “theories”) Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 16 is largely absent in such computational frameworks. The intensive to implement with graph-like representations), and, as possibility of representing and handling, in an integrated way, a consequence, the typology of encoded knowledge is biased an heterogeneous amount of common sense conceptual towards the ``classical" (but unsatisfactory) representation of representations (and the related reasoning mechanisms), in fact, concepts in terms of necessary and sufficient conditions [10]. is not sufficiently addressed both by the symbolic-based This characterization, however, is problematic for modelling “chunk-structures” adopted by the most common general CAs real world concepts and, on the other hand, the so called (e.g. SOAR) and by fully connectionist architectures (e.g. common-sense knowledge components (i.e. those that, allow to LEABRA). This aspect is problematic also in the hybrid characterize and process conceptual information in terms of solutions adopted by CAs such as CLARION [11] or ACT-R typicality and involving, for example, prototypical and (the different reasons leading to a non satisfactory treatment of exemplar based representations and reasoning mechanisms) is this aspect are detailed in [12]). This type of knowledge, largely absent. This problem arises despite the fact that the however, is exactly the type of “cognitive information” chunks in SOAR can be represented as a sort of frame-like crucially used by humans for heuristic reasoning and decision structures containing some common-sense (e.g. prototypical) making. This paper presents an analysis of the current situation information [12]. W.r.t. to the size problem, the SOAR by proposing a comparison of the representational level of knowledge level is also problematic. SOAR agents, in fact, are SOAR, ACT-R, CLARION and Vector-LIDA. Finally, we not endowed with general knowledge and only process ad-hoc suggest that a possible way out to deal with this problem could built (or task-specific learned) symbolic knowledge structures. be represented by the integration of external cognitive systems into the knowledge representation and processing mechanisms B. ACT-R of general cognitive architectures. Some initial efforts in this direction, have been proposed (see e.g. [13, 14]) and will be ACT-R is a cognitive architectures explicitly inspired by presented and discussed. theories and experimental results coming from human cognition. Here the cognitive mechanisms concerning the III. KNOWLEDGE REPRESENTATION IN CAS knowledge level emerge from the interaction of two types of In the following we provide a short overview of: SOAR [3], knowledge: declarative knowledge, that encodes explicit facts ACT-R [2], CLARION [11] and LIDA [15] (in its novel that the system knows, and procedural knowledge, that encodes version known as Vector-LIDA [16]). The choice of these rules for processing declarative knowledge. In particular, the architecture has been based on the fact that they represent some declarative module is used to store and retrieve pieces of of the most widely used systems (adopted in scenarios ranging information (called chunks, featured by a type and a set of from robotics to video-games) and their representational attribute-value pairs, similar to frame slots) in the declarative structures present some relevant differentiations that are memory. ACT-R employs a wide range of sub-symbolic interesting to investigate in the light of the issues raised in this processes for the activation of symbolic conceptual chunks paper. By analyzing, in brief, such architectures we will representing the encoded knowledge. Finally, the central exclusively focus on the description of their representational production system connects these modules by using a set of IF- frameworks since a more comprehensive review of their whole THEN production rules using a set of IF-THEN production mechanisms is out of the scope of the present contribution rules. Differently from SOAR, ACT-R allows to represent the (detailed reviews of their mechanisms are described in [17]; information in terms of prototypes and exemplars and allow to and [18]). We will show how all of them are affected, at perform, selectively, either prototype or exemplar-based different levels of granularity, by both the size and the categorization. This means that this architecture allows the knowledge homogeneity problems. modeller to manually specify which kind of categorization strategy to employ according to his specific needs. Such A. SOAR architecture, however, only partially addresses the homogeneity problem since it does not allow to represent, SOAR is one of the oldest cognitive architectures. This jointly, these different types of common-sense representations system was considered by Newell a candidate for a Unified for the same conceptual entity (i.e. it does not assume a Theory of Cognition [19]. One of the main themes in SOAR is heterogeneous perspective). As a consequence, it is also not that all cognitive tasks can be represented by problem spaces able to autonomously decide which of the corresponding that are searched by production rules grouped into operators. reasoning procedures to activate (e.g. prototypes or exemplars) These production rules are red in parallel to produce reasoning and to provide a framework able to manage the interaction of cycles. From a representational perspective, SOAR exploits such different reasoning strategies (however its overall symbolic representations of knowledge (called chunks) and use architectural environment provides, at least in principle, the pattern matching to select relevant knowledge elements. possibility of implementing cascade reasoning processes Basically, where a production match the contents of declarative triggering one another). Even if, in such architecture, some (working) memory the rule fires and then the content from the attempts exist concerning the design of harmonization declarative memory (called Semantic Memory in SOAR) is strategies between different types of common-sense conceptual retrieved. This system adheres strictly to the Newell and categorizations (e.g. exemplars-based and rule based, see [20]) Simon's physical symbol system hypothesis which assumes however they do not handle the problem concerning the that symbolic processing is a necessary and sufficient condition interaction of the prototype or exemplars-based processes for intelligent behavior. The SOAR system encounter, in according to the results coming from the experimental general, the standard problems affecting symbolic formalisms cognitive science (for example: the old item effect, privileging at the representational level: it is not well equipped to deal with exemplars w.r.t. prototypes is not modelled. See again [12] for common-sense knowledge representation and reasoning (since a detailed analysis of this aspect). Summing up: w.r.t. the approximate comparisons are hard and computationally knowledge homogeneity problem, the components needed to Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 17 fully reconcile the Heterogeneity approach with ACT-R are hierarchical structures in vectors or connectionist models. present, however they have not been fully exploited yet. These reduced description vectors can be expanded to obtain Regarding the size problem: as for SOAR, ACT-R agents are the whole structure, and can be used directly for complex usually equipped with task-specic knowledge and not with calculations and procedures, such as making analogies, logical general cross-domain knowledge. In this respect some relevant inference, or structural comparison. Vectors in this framework attempts to overcome this limitation have been recently done are treated as symbol-like representations, thus enabling by extending the Declarative Memory of the architecture. They different kind of operations executed on them (e.g. simple will be discussed in section E along with their current forms of compositionality via vectors blending). Vector- LIDA, implications. encounters the same limitations of the other CAs since i) its agents are not equipped with a general cross-domain C. CLARION knowledge and therefore can be only used in very narrow tasks (their knowledge structure is either ad hoc build or ad hoc CLARION is a hybrid cognitive architecture based on the learned). Additionally, this architecture does not address the dual-process theory of mind. From a representational problem concerning the heterogeneity of the knowledge perspective, processes are mainly subject to the activity of two typologies. In particular its knowledge level does not represent sub-systems, the Action Centered Sub-system (ACS) and the the common-sense knowledge components such as prototypes Non-Action Centered Sub-system (NACS). Both sub-systems and exemplars (and the related reasoning strategies). In fact, as store information using a two-layered architecture, i.e., they for CLARION, despite vector-representations allow to perform both include an explicit and an implicit level of many kind of approximate comparisons and similarity-based representation. Each top-level chunk node is represented by a reasoning (e.g. in tasks such as categorization), the peculiarity set of (micro)features in the bottom level (i.e., a distributed concerning prototype or exemplars based representations representation). The (micro)features (in the bottom level) are (along with the the design of the interaction between their connected to the chunk nodes (in the top level) so that they can different reasoning strategies) are not provided. In this respect, be activated together through bottom-up or top-down however an element that is worth-noting is represented by the activation. Therefore, in general, a chunk is represented by fact that the Vector-LIDA representational structures are very both levels: using a chunk node at the top level and distributed close to the framework of Conceptual Spaces. Conceptual feature representation at the bottom level. W.r.t. to the Spaces are a geometric knowledge representation framework knowledge size and homogeneity problems, CLARION, proposed by Peter Gärdenfors [21]. They can be thought as a encounter problems with both these aspects since i) there are particular class of vector representations where knowledge is no available attempts aiming at endowing such architecture represented as a set of quality dimensions, and where a with a general and cross-domain knowledge ii) the dual-layered geometrical structure is associated to each quality dimension. conceptual information does not provide the possibility of They are discussed in more detail in section 5. The encoding (manually or automatically via learning cycles) the convergence of the Vector-LIDA representation towards information in terms of the heterogeneous classes of Conceptual Spaces could enable, in such architecture, the representations presented in the section 2. In particular: the possibility of dealing with at least prototype and exemplars- main problematic aspect concerns the representation of the based representations and reasoning, thus overcoming the common-sense knowledge components. As for SOAR and knowledge homogeneity problem. ACT-R, also in CLARION the possible co-existence of typical representations in terms of prototypes, exemplars and theories E. Attempts to Overcome the Knowledge Limits (and the interaction among them) is not treated. In terms of As mentioned above, some initial efforts to deal with the reasoning strategies, notwithstanding that the implicit limited knowledge availability for agents endowed with knowledge layer based on neural network representations can cognitive architecture have been done. In particular, within provide forms of non monotonic reasoning (e.g. based on Mind'sEye program (a DARPA founded project), the similarity), such kind of similarity-based reasoning is currently knowledge layers of ACT-R architecture have been not grounded on the mechanisms guiding the decision choices semantically extended with an external ontological content followed, for example, by prototype or exemplars-based coming from three integrated semantic resources composed by reasoning. the lexical databases WordNet [22], FrameNet [23] and by a branch of the top level ontology DOLCE [24] related to the D. Vector-LIDA event modelling. In this case, the amount of semantic Vector LIDA is a cognitive architecture employing, at the knowledge selected for the realization of the Cognitive Engine representational level, high-dimensional vectors and reduced (one of the systems developed within the MindEye Program) descriptions. High-dimensional vector spaces have interesting and for its evaluation, despite by far larger w.r.t. the standard properties that make them attractive for representations in ad-hoc solutions, was tailored on the specific needs of the cognitive models. The distribution of the distances between system itself. It, in fact, was aimed at solving a precise task of vectors in these spaces, and the huge number of possible event recognition trough a video-surveillance intelligent vectors, allow noise-robust representations where the distance machinery; therefore only the ontological knowledge about the between vectors can be used to measure the similarity (or events was selectively embedded in it. While this is a dissimilarity) of the concepts they represent. Moreover, these reasonable approach in an applicative context, still does not high-dimensional vectors can be used to represent complex allow to test the general cognitive mechanisms of a Cognitive structures, where each vector denotes an element in the Architecture on a general, multi faceted and multi-domain, structure. However, a single vector can also represent one of knowledge. Therefore it does not allow to evaluate strictu these same complex structures in its entirety by implementing a sensu to what extent the designed heuristics allowing to reduced description, a mechanism to encode complex retrieve and process, from a massive and composite knowledge Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 18 base, conceptual knowledge can be considered satisfyicing The knowledge level of DUAL PECCS is heterogeneous in w.r.t. the human performances. More recent works have tried to nature since it is explicitly based and designed on the completely overcome at least the size problem of the assumption that concepts are “heterogeneous proxytypes” [27] knowledge level. To this class of works belongs that one and, as such, they are composed by heterogeneous knowledge proposed by Salvucci [9] aiming at enriching the knowledge components selectively and contextually activated in working model of the Declarative Memory of ACT-R with a world-level memory. In particular, by following the proposal presented in knowledge base such as DBpedia (i.e. the semantic version of [28, 29], the representational level of DUAL PECCS couples Wikipedia represented in terms of ontological formalisms) and Conceptual Spaces representations and ontological knowledge a previous one proposed in [25] presenting an integration of the (consisting in the Cyc ontology) for the same conceptual entity. ACT-R Declarative and Procedural Memory with the Cyc Conceptual Spaces [21] is used to represent and process the ontology [26] (one of the widest ontological resources common-sense conceptual information. In such framework, to currently available containing more than 230,000 concepts). each quality dimension is associated a geometrical (topological Both the wide-coverage integrated ontological resources, or metrical) structure. In some cases, such dimensions can be however, represents conceptual information in terms of directly related to perceptual mechanisms; examples of this symbolic structures and encounter the standard problems kind are temperature, weight, brightness, pitch. In other cases, affecting this class of formalisms and discussed above. Some dimensions can be more abstract in nature. In this setting, of these limitations can be, in principle, partially overcome by concepts correspond to convex regions, and regions with such works, since the integration of such wide-coverage different geometrical properties correspond to different sorts of ontological knowledge bases with the ACT-R Declarative concepts [21]. Here, prototypes and prototypical reasoning Memory allows to preserve the possibility of using the have a natural geometrical interpretation: prototypes common-sense conceptual processing mechanisms available in correspond to the geometrical centre of a convex region (the that architecture (e.g. prototype and exemplars based). centroid). Also exemplars-based representation can be Therefore, in principle, dealing with the size problem also represented as points in a multidimensional space, and their allows to address some aspects concerning the heterogeneity similarity can be computed as the intervening distance between problem. Still, however, remains the problem concerning the each two points, based on some suitable metrics (such as lack of the representation of common-sense information to Euclidean and Manhattan distance etc.). The ontological which such common-sense architectural processes can be component, on the other hand, is used to provide and process applied: e.g. a conceptual retrieval based on prototypical traits the “classical” knowledge component for the same conceptual (i.e. a prototype-based categorization) cannot be performed on entity. such integrated ontological knowledge bases since these symbolic systems do not represent at all the typical information The representational level of DUAL PECCS (and the associated to a given concept ([12] presents an experiment on corresponding knowledge processing mechanisms) has been this aspect). In addition, as already mentioned, it remains not successfully integrated with the representational counterpart of yet addressed the problem concerning the interaction, in a some available CAs [14, 30] by extending, de facto, the general and principled way, of the different types of common- knowledge representation and processing capabilities of sense processes involving different representations of the same cognitive architectures based on diverse representational conceptual entity. In the light of the arguments presented above assumptions. One of the main novelties introduced by DUAL it can be argued, therefore, that the current proposed solutions PECCS (and therefore one of the main advantages obtained by for dealing with the knowledge problems in CAs are not the CAs extended with such external cognitive system) consists completely satisfactory. In particular, the integrations with in the fact that it is explicitly designed the flow of interaction huge world-level ontological knowledge bases can be between common-sense categorization processes (based on considered a necessary solution for solving size problem. It is, prototypes and exemplars and operating on conceptual spaces however, insufficient for dealing with the knowledge representations) and the standard deductive processes homogeneity problem and with the integration of the common- (operating on the ontological conceptual component). The sense conceptual mechanisms activated on heterogeneous harmonization regarding such different classes of mechanisms bodies of knowledge, as assumed in the heterogeneous has been devised based on the tenets coming from the dual representational perspective. In the next sections we outline a process theory of reasoning [31, 32]. Additionally, in DUAL possible alternative solution that, despite being not yet fully PECCS, also the interaction of the categorization processes developed is, in perspective, suitable to account for both for the occurring within the class of non monotonic categorization heterogeneous aspects in conceptualization and for the size mechanisms (i.e. prototypes and exemplars-based problems. categorization) has been devised and is dealt with at the Conceptual Spaces level. This latter aspect is of particular IV. INTEGRATING EXTERNAL COGNITIVE SYSTEMS IN CA interest in the light of the multifaceted problem concerning the heterogeneity of the encoded knowledge. In fact, since the Recently some available conceptual categorization systems, design of the interaction of the the different processes operating explicitly assuming the heterogeneous representational with heterogeneous representations still represents, as seen hypothesis and integrated with wide-coverage knowledge bases before, a largely unaddressed problem in current CAs, this (such as Cyc) have been developed and integrated with the system shows the relative easiness that its knowledge knowledge level of available CAs. For our purposes, we will framework (and, in particular, the Conceptual Spaces consider here the DUAL PECCS system [13, 14]. We will not component) provides to naturally model the dynamics between discuss the results obtained by such system in tasks of prototype and exemplars-based processes. For what concerns conceptual categorization, since they have been already the size problem, finally, the possibile grounding of the presented elsewhere [14]. We shall briefly focus, in the Conceptual Spaces representational component with symbolic following, on the representational level of the system. structures enables the integration with wide-coverage Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 19 knowledge bases such as Cyc. Thus, the solution adopted in 17. Vernon, D, Metta, G., Sandini, G,., A survey of artificial cognitive DUAL PECCS is, in principle, able to deal with both the size systems:Implications for the autonomous development of mental capabilities in computational agents, IEEE Transactions on Evolutionary and the knowledge homogeneity problems affecting the CAs. Computation 11 (2), 2007. In particular, the extension of the Declarative Memories of the 18. Langley, P., Laird, J., Rogers, S., Cognitive architectures: Research current CAs with this external cognitive system allowed to issues and challenges, Cognitive Systems Research 10 (2), 141-160, empower the knowledge processing and categorization 2009. capabilities of such general architectures (an important role, in 19. Newell, A. Unified theories of cognition. Cambridge, MA: Harvard this respect, is played by the Conceptual Spaces component). University Press, 1990. Despite there is still room of improvements and further 20. Anderson, J., Betz, J., A hybrid model of categorization, Psychonomic investigations, this seems a promising way to deal with the Bulletin & Review 8 (4) 629-647, 2001. both the knowledge problems discussed in this paper. 21. Gärdenfors, Conceptual spaces: The geometry of thought, MIT press, 2000. ACKNOWLEDGMENTS 22. Fellbaum, C. (ed.): WordNet - An Electronic Lexical Database. I thank the reviewers of the EuCognition Conference for Cambridge, Massachusetts. MIT Press, 1998. their useful comments. The arguments presented in this paper 23. Fillmore, C.J.,The case for case. Bach, E., Harms, T. eds. Universals in have been discussed in different occasions with Christian Linguistic Theory. New York: Rinehart and Wiston, 1968. Lebiere, Alessandro Oltramari, Antonio Chella, Marcello 24. Masolo, C., Borgo, S., Gangemi, A., Guarino, N., & Oltramari, A., Frixione, Peter Gärdenfors, Valentina Rho and Daniele Wonderweb deliverable d18, ontology library (final). ICT project, Radicioni. I would like to thank them for their feedback. 33052, 2003. 25. Ball, J., Rodgers, S., Gluck, K., Integrating act-r and cyc in a large-scale REFERENCES model of language comprehension for use in intelligent agents, in: AAAI workshop, pp. 19-25, 2004. 26. Lenat, D., Cyc: A large-scale investment in knowledge infrastructure, 1. Oltramari A., Lebiere C., Pursuing Artificial General Intelligence By Communications of the ACM 38 (11), 33-38, 1995. Leveraging the Knowledge Capabilities Of ACT-R, AGI 2012 (5th International Conference on "Artificial General Intelligence”), Oxford, 27. Lieto A., A Computational Framework for Concept Representation in 2012. Cognitive Systems and Architectures: Concepts as Heterogeneous Proxytypes, Procedia Computer Science, 41, 6–14, http://dx.doi.org/ 2. Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S., Lebiere, C., & 10.1016/j.procs.2014.11.078, 2014. Qin, Y., An integrated theory of the mind, Psychological Review, 111(4), 1036-1060, http://dx.doi.org/10.1037/0033-295X.111.4.1036, 2004. 28. Frixione M, Lieto A., Towards an Extended Model of Conceptual Representations in Formal Ontologies: A Typicality-Based Proposal, 3. Laird, John E., The Soar Cognitive Architecture, MIT Press, 2012. Journal of Universal Computer Science 20 (3) (2014) 257–276, 2014. 4. Newell, A. The knowledge level, Artificial intelligence 18 (1), 87-127, 29. Lieto, A., Chella, A., Frixione, M., Conceptual Spaces for Cognitive 1982. Architectures: A Lingua Franca for Different Levels of Representation, 5. Gigerenzer, G., Todd, P., Simple Heuristics that make us smart”, Oxford In Biologically Inspired Cognitive Architectures, 19 (2), 1-9, 2017. University Press, 1999. 30. Lieto, A., Radicioni, D.P., Rho, V., Mensa, E., Towards a Unifying 6. Simon, H., A Behavioral Model of Rational Choice, in Mathematical Framework for Conceptual Representation and Reasoning in Cognitive Essays on Rational Human Behaviour in Social Setting. NY: Wiley, Systems, in Intelligenza Artificiale, forthcoming. 1957. 31. Evans, J., Frankish, K., In two minds: Dual processes and beyond, 7. Minkowski M., Explaining the Computational Mind, MIT Press, 2013. Oxford University Press, 2009. 8. Lieto, A., Radicioni, D.P. From Human to Artificial Cognition and back: 32. Stanovich, K., West, R., Advancing the rationality debate, Behavioral Challenges and Perspectives of Cognitively-Inspired AI sistems, and brain sciences, 23 (05), 701-717, 2000. Cognitive Systems Research, 39 (2), pp. 1-3, 2016. 9. Salvucci., D., Endowing a Cognitive Architecture with World Knowledge, Proceedings of the 36th Annual Meeting of the CogSci Soc., 2014. 10. Frixione, M, Lieto A., Representing concepts in formal ontologies: Compositionality vs. typicality effects, Logic and Logical Philosophy 21 (4) (2012) 391–414, 2012. 11. Sun, R. The CLARION cognitive architecture: Extending cognitive modeling to social simulation. Cognition and multi-agent interaction pp. 79-99, 2006. 12. Lieto, A., Lebiere C., Oltramari, A. The Knowledge Level in Cognitive Architectures: Current Limitations and Possibile Developments, Cognitive Systems Research, Cognitive Systems Research, forthcoming. 13. Lieto, A., Radicioni, D.P., Rho, V., A Common Sense Conceptual Categorization System Integrating Heterogeneous Proxytypes and the Dual Process of Reasoning. In Proc. of IJCAI 2015, AAAI Press, 2015. 14. Lieto, A., D.P. Radicioni, V. Rho, Dual PECCS: A Cognitive System for Conceptual Representation and Categorization Journal of Experimental and Theoretical Artificial Intelligence, Taylor and Francis. doi: http:// dx.doi.org/10.1080/0952813X.2016.1198934, 29 (2), 2017. 15. Franklin, S., F. Patterson, F., The Lida architecture: Adding new modes of learning to an intelligent, autonomous, software agent, 764-1004, 2006. 16. Snaider, S. Franklin, Vector Lida, Procedia Computer Science 4,188-203, 2014. Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 20