Chapter proposal for 'Visions of Mind-Architectures for Cognition and Affect'. The CHREST .... with Piaget's theory of development (Piaget, 1970). Future Trends.
Chapter proposal for ‘Visions of Mind-Architectures for Cognition and Affect’ The CHREST architecture of cognition: Listening to empirical data Fernand Gobet (University of Nottingham) and Peter C.R. Lane (University of Hertfordshire) Introduction A number of approaches have been taken in attempting to understand the human mind, spanning philosophy, psychology, neuroscience, and computer science. An influential line of research has been to develop computational architectures that closely simulate human behaviour in a variety of domains; examples of this approach include ACT-R (Anderson & Lebière, 1998), Soar (Newell, 1990), and EPAM (Feigenbaum & Simon, 1984). More recently, the computational architecture CHREST (Chunk Hierarchy and REtrieval Structures) (Gobet et al., 2001; Gobet & Simon, 2000; Lane, Cheng, & Gobet, 2000) has simulated data in a number of domains, including expert behaviour, verbal learning, first language acquisition, and implicit learning. The aim of this chapter is to provide an introduction to CHREST, to illustrate the kind of data that have been successfully simulated, and to show what insight CHREST offers on some of the difficult questions facing researchers studying the mind, such as intuition, consciousness, implicit learning, and the link between emotion and cognition. Background CHREST views the mind as a collection of emergent properties produced by the interaction of short-term memory, long-term memory, learning, perception, and decisionmaking structures and processes. In particular, it is assumed that there is a close interaction between perception, learning, and memory; for example, what is known by the system will direct attention and perception (through the eye movements), and, in turn, perception will direct what will be learned next. Another important characteristic of the
architecture is that it is constrained by a number of parameters, such as the capacity of short-term memory, the rate with which new elements can be learned, and the time to transfer information from long-term memory to short-term memory. Thus, a theme central to this approach is that the human cognitive system satisfies Simon’s requirements of bounded rationality (Simon, 1969). Finally, all operations carried out by the system have time costs, which makes it possible to provide detailed simulations of empirical data. The general approach defended in this chapter is that, in order to develop a theory of the mind, one has to use empirical data to constrain the number of possible architectures (see also, Newell, 1990). Conversely, the architecture makes new empirical predictions that can be tested and that can be used to further develop it. Indeed, CHREST has already made new predictions in the field of expert behaviour, predictions that were later supported by empirical data (Gobet & Waters, in press). Domains already successfully tackled by CHREST As proposed above, we claim that a strong approach to develop a computational theory of mind is to compare the predictions of the architecture with the empirical data. How does CHREST fare in this respect? CHREST was originally developed to explain expert behaviour in chess, a domain that draws upon a variety of cognitive abilities, such as perception, memory, decision making, and problem solving. It is able to simulate a number of phenomena closely, such as the novices’ and chess masters’ eye movements, performance in memory-recall experiments (including errors and the detail of the piece placements), and how aspects of look-ahead search evolve as a function of skill. CHREST has also been applied to other domains of expertise, such as memory for computer programs and the acquisition of multiple representations in physics. Beyond expertise, our architecture has explained phenomena in verbal learning, the acquisition of language (both the acquisition of syntax and vocabulary), developmental psychology, implicit learning, and concept formation (see Gobet et al., 2001, for pointers to the literature). It is currently being applied to low-level vision and to explaining how
low-level and high-level knowledge interact to produce intelligent behaviour. While the range of domains covered is less than that of architectures such as ACT-R or Soar, it should be pointed out that CHREST incorporates many fewer degrees of freedom than these theories. Developing a complete mind: Some hard questions While CHREST is far from implementing a complete mind, we can be confident that the components in place are at least sufficient to produce human-like behaviour in a number of domains—sometimes with a remarkable degree of detail. In this section, we will speculate about how CHREST can help explain some difficult questions in cognitive science. For example, we will show that the limits inherent in the architecture, as well as its ability to acquire aspects of language, may be used to develop a theory of consciousness that can make testable predictions; that intuition may be the product of pattern-recognition mechanisms, as implemented in the architecture (see also (Simon, 1979)); that emotions and motivation may mediate some of the parameters of the architectures, such as the time to create a new long-term memory structure or to add information to an extant structure; and that the architecture shares interesting features with Piaget’s theory of development (Piaget, 1970). Future Trends While we have focused upon the psychological aspects of CHREST, it must be realised that there are also potential technical applications in the area of AI. For example, the architecture’s ability to seamlessly combine low-level and high-level aspects of cognition can be used in image analysis. A reasonable criticism against our approach is that it mainly uses symbolic (although domain-representative) input to train the architecture. We are currently working to use less abstract types of input, such as bitmap images; we are also aiming at ‘embodying’ CHREST by connecting it to a mobile robot. Although we are obviously biased towards our own approach, we also believe that true progress in understanding the human mind will be made by comparing cognitive
architectures, both those closely simulating empirical data and those developed based more on basic architectural principles, such as the necessity of parallelism or the need to minimize computation. An interesting question is whether the conclusions reached by simulating the detail of human behaviour will coincide with conclusions based upon architectural considerations. Conclusion The achievements of CHREST in explaining a variety of empirical phenomena entitle us to suggest some ideas about what a complete model of the mind would look like. Interaction between perception and cognition, the capability to incrementally acquire a vast and integrated structure of nodes and links, as well as (paradoxically) strong architectural limits are the key features behind CHREST’s success in simulating a variety of data, and, we suggest, behind any successful computational theory of mind. While carrying out detailed simulation of human behaviour is time-consuming and difficult, it may be the most efficient way to search the space of possible architectures for a mind. References Anderson, J. R., & Lebière, C. (Eds.). (1998). The atomic components of thought. Mahwah, NJ: Erlbaum. Feigenbaum, E. A., & Simon, H. A. (1984). EPAM-like models of recognition and learning. Cognitive Science, 8, 305-336. Gobet, F., Lane, P. C. R., Croker, S., Cheng, P. C.-H., Jones, G., Oliver, I., & Pine, J. (2001). Chunking mechanisms in human learning. Trends in Cognitive Sciences, 5(6), 236-243. Gobet, F., & Simon, H. A. (2000). Five seconds or sixty? Presentation time in expert memory. Cognitive Science, 24, 651-682. Gobet, F., & Waters, A. J. (in press). The role of constraints in expert memory. Journal of Experimental Psychology: Learning, Memory & Cognition. Lane, P. C. R., Cheng, P. C.-H., & Gobet, F. (2000). CHREST+: Investigating how humans learn to solve problems using diagrams. AISB Quarterly, 103, 24-30. Newell, A. (1990). Unified theories of cognition. Cambridge, MA: Harvard University Press. Piaget, J. (1970). L'épistémologie génétique. Paris: PUF. Simon, H. A. (1969). The sciences of the artificial. Cambridge: MIT Press. Simon, H. A. (1979). Models of thought (Vol. 1). New Haven: Yale University Press.