Cognitive Architectures: Where do...
Cognitive Architectures: Where do we go from here? W��odzis��aw DUCHa,1, Richard J. OENTARYOb, and Michel PASQUIERb aDept. of Informatics, Nicolaus Copernicus University, Toru��, Poland bSchool of Computer Engineering, Nanyang Technological University, Singapore Abstract. Cognitive architectures play a vital role in providing blueprints for building future intelligent systems supporting a broad range of capabilities similar to those of humans. How useful are existing architectures for creating artificial general intelligence? A critical survey of the state of the art in cognitive architectures is presented providing a useful insight into the possible frameworks for general intelligence. Grand challenges and an outline of the most promising future directions are described. Keywords: cognitive architectures, artificial general intelligence, neurocognitive models, intelligent agents. 1. Introduction A long-term goal for artificial general intelligence (AGI) is to create systems that will exceed human level of competence in a large number of areas. There is a steady progress toward this goal in several domains, including recognition of specific patterns, memorization and retrieval of vast amount of information, interpreting signals and other types of numerical information, autonomous control, board games and reasoning in restricted domains. Yet even in lower level cognitive functions, such as object recognition or scene analysis artificial systems are still far behind the natural ones. Higher-level cognitive functions, such as language, reasoning, problem solving or planning, involve complex knowledge structures and are much more difficult to realize. Various types of memory stored by the brain facilitate recognition, association, semantic interpretation and rapid retrieval of large amounts of complex information patterns. At quite basic level organization of storage and retrieval of information in computers is completely different than in brains. Computing architectures are universal only in principle, in practice they always constrain information processing in specific ways. Computers are better in many tasks than brains, and vice versa, brains are better in many important tasks then computers. It is not clear at all whether cognitive architectures (CA) running on conventional computers, will reach the flexibility of the brain in lower or higher-level cognitive functions. Traditionally higher cognitive functions, such as thinking, reasoning, planning, problem solving or linguistic competencies have been the focus of artificial intelligence (AI), relaying on symbolic problem solving to build complex knowledge structure. These functions involve sequential search processes , while lower cognitive functions, such as perception, motor control, sensorimotor actions, associative memory recall or categorization, are accomplished on a faster time scale in a parallel way, without stepwise deliberation. Embodiment is a powerful trend in robotics and there is now a general agreement that the meaning of many concepts should be grounded in embodied, sensorimotor representations. While symbolic approximations that account for sensorimotor processes are certainly possible not much is known about their limitations, for example how deep the grounding of symbols should be, and how to achieve it through embodied cognition. Perhaps the dream of creating a General Problem Solver  may be realized with relatively minor extensions to symbolic cognitive architectures, while detailed understanding of animal behavior and creating flexible mobile robotic applications may require a distributed approach to embodied cognition. Analysis of existing cognitive architectures should facilitate understanding of limitations of different approaches. Many general ideas seem to explain everything but do not scale up well to real applications, therefore a clear notion what exactly AGI should do is necessary. 1 Corresponding author, Dept. of Informatics, Nicolaus Copernicus Uni., Grudzi��dzka 5, Toru��, Poland, Google: ���W. Duch���.
2. Grand challenges for AGI What should be required from an AI system to be worthy of the ���Artificial General Intelligence��� name? Artificial Intelligence has focused on many specific approaches to problem solving, useful for development of expert systems, neglecting its initial ambitious goals. One requirement for AGI, storing and manipulation of vast amount of knowledge, has been addressed by the Cyc project . Started in 1984 a huge frame-based knowledge base has been constructed, but the list of its ���potential applications��� has not been replaced by real applications for decades. Perhaps the biggest mismatch between AI reality and popular expectations is in the language-related domains, for example in general purpose conversational systems, developed mostly in the form of various chatterbots by commercial companies and enthusiastic individuals. Restricted form of the Turing test  (the full test being too difficult to try), called Loebner Prize competition2, has been won for almost two decades by chatterbots based on old template matching techniques, or more recently contextual pattern matching techniques. Such programs have no chance to develop real understanding of language and use it in meaningful dialogs or texts analysis, but may be used for stereotyped question/answer systems or ���impersonation���. Carpenter and Freeman have proposed a ���personal Turing test��� , where a person tries to guess if the conversation is done with a program or a real personally known individual. Human behavior includes the ability to impersonate other people, and the personal Turing test may be an interesting landmark step on the road to general intelligence. Another area that poses remarkable challenge to AI is word games, and in particular the 20-questions game. Word games require extensive knowledge about objects and their properties, but not about complex relations between objects. Different methods of knowledge representation may be used in different applications, from quite simple, facilitating efficient use of knowledge, to quite involved, needed only in deep reasoning. In fact simple vector-space techniques for knowledge representation are sufficient to play the 20- question game . Success in learning language depends on automatic creation and maintenance of large- scale knowledge bases, bootstraping on the resources from the Internet. Question/answer systems pose even more demanding challenge, and in this area a series of competitions organized at Text Retrieval Conference (TREC) series may be used to measure progress. Intelligent tutoring systems are the next great challenge, but there seem to be no clearly defined milestones in this field. Feigenbaum  proposed as a grand challenge building a super-expert system in a narrow domain. This seems to go in a direction of specialized, rather than general intelligence, but one may argue that a super- expert without general intelligence needed for communication with humans is not feasible. Sophisticated reasoning by human experts and artificial systems in such fields as mathematics, bioscience or law may be compared by a panel of experts who will pose problems, rise questions, and ask for further explanations to probe the understanding of the subject. A good example of such challenge is provided by the Automated Theorem Proving (ATM) System Competitions (CASC) in many sub-categories. An interesting step toward general AI in mathematics would be to create general theorem provers, perhaps using meta-learning techniques that rely on specialized modules. Automatic curation of genomic/pathways databases and creation of models of genetic and metabolic processes for various organisms poses great challenges for super-experts, as the amount of information in such databases exceeds by far human capacity to handle it. Defining similar challenges and milestones towards AGI in other fields is certainly worthwhile. The ultimate goal would be to develop programs that will advice human experts in their work, evaluating their reasoning, perhaps even adding some creative ideas. DARPA in the ���Personal Assistants that Learn��� (PAL) program sponsors a large-scale effort in similar direction. Nilsson  has argued for development of general purpose educable systems that can be taught skills needed to perform human jobs, and to measure which fraction of these jobs can be done by AI systems. Building one such system replaces the need for building many specialized systems, as already Allan Turing  has noted proposing a ���child machine��� in his classical paper. Some human jobs are knowledge-based and can be done by information processing systems, where progress may be measured by passing a series of examinations, as is done in such fields as accounting. However, most human jobs involve manual labor, requiring senso-motoric coordination that should be mastered by household robots or autonomous vehicles. The DARPA Urban Challenge competition (2007) required integration of computer vision, signal processing, control and some reasoning. It is still simpler than control of a humanoid robot, where direct interaction of robots with people will require an understanding of perception, controlling of attention, learning casual models from observations, and hierarchical learning with different temporal scales. Creation of partners or personal assistants, rather than complete replacements for 2 See http://www.loebner.net for information on the Loebner competition.
human workers, may be treated as a partial success. Unfortunately specific milestones for this type of applications have yet to be precisely defined. Some ordering of different jobs from the point of view of difficulty to learn them could be worthwhile. In fact many jobs have already been completely automatized, reducing the number of people in manufacturing, financial services, printing houses etc. In most cases alternative organization of work is to be credited for reduction in the number of jobs (plant and factory automation, ATM machines, vending machines), not because of deployment of AI systems. A detailed roadmap to AGI should thus be based on detailed analysis of the challenges, relationships between various functions that should be implemented to address them, system requirements to achieve these functions and classes of problems that should be solved at a given stage. 3. Cognitive architectures Cognitive architectures are frequently created to model human performance in multimodal multiple task situations  rather than to create AGI. A short critical review of selected cognitive architectures that can contribute to development of AGI is provided below. Allen Newell in his 1990 book Unified Theories of Cognition  provided 12 criteria for evaluation of cognitive systems: adaptive behavior, dynamic behavior, flexible behavior, development, evolution, learning, knowledge integration, vast knowledge base, natural language, real-time performance, and brain realization. These criteria have been analyzed and applied to ACT-R, Soar and classical connectionist architectures  but such fine-grained categorization makes comparison of different systems rather difficult. Without going into such details we shall propose below a simpler taxonomy, give some examples of different types of cognitive systems that are currently under development, and provide a critique and some recommendations for better systems. Surveys on the system organization and working mechanisms of a few cognitive architectures that have already been published  were not written from the AGI point of view. Two key design properties that underlie the development of any cognitive architecture are memory and learning. The importance of memory has been stressed from different perspectives in a few recent books - . Various types of memory serve as a repository for background knowledge about the world and oneself, about the current episode of activity, while learning is the main process that shapes this knowledge. Together learning and memory form the rudimentary aspects of cognition on which higher-order functions and intelligent capabilities, such as deliberative reasoning, planning, and self-regulation, are built. Organization of memory depends on the knowledge representation schemes. A simple taxonomy of cognitive architectures based on these two main features leads to a division of different approaches into three main groups (Fig. 1): symbolic, emergent, and hybrid models. Roughly speaking symbolic architectures focus on information processing using high-level symbols or declarative knowledge, in a classical AI top-down, analytic approach. Emergent architectures use low-level activation signals flowing through a network consisting of numerous processing units, a bottom-up process Symbolic Emergent Hybrid Cognitive architectures Memory ��� Rule-based memory ��� Graph-based memory Learning ��� Inductive learning ��� Analytical learning Memory ��� Globalist memory ��� Localist memory Learning ��� Associative learning ��� Competitive learning Memory ��� Localist-distributed ��� Symbolic-connectionist Learning ��� Bottom-up learning ��� Top-down learning Fig. 1. Simplified taxonomy of cognitive architectures