MALTA: Enhancing ACT-R with a Holographic Persistent Knowledge Store

MALTA: Enhancing ACT—R with a Holographic Persistent Knowledge Store Matthew F. Rutledge-Taylor (mrtaylo2@connect.carleton.ca) Institute of Cognitive Science, Carleton University, 1125 Colonel By Drive Ottawa, Ontario, K1S 5B6 Canada Robert L. West (robert_west@carleton.ca) Institute of Cognitive Science, Department of Psychology, Carleton University, 1125 Colonel By Drive Ottawa, Ontario, K1S 5B6 Canada Abstract There is a concern that ACT—R (Anderson & Lebiere, 1998), and other cognitive architectures do not adequately account for long-term memory (Schultheis, Barkowsky, & Bertel, 2006). These architectures, while ideal for modeling single tasks, do not address the issue of preserving a task neutral knowledge store, which persists between tasks (Rutledge- Taylor, 2005). A dynamical holographic associative memory system, DSHM, based on Jones and Mewhort’s model of the lexicon, BEAGLE (2007), is presented. A suggestion for a new cognitive architecture, MALTA, which combines elements of ACT—R and DSHM is described. Arguments for why this new architecture addresses some concerns with ACT-R’s declarative memory system are offered. Keywords: cognitive modeling; cognitive science; ACT—R; knowledge representation; holographic associative memory. Introduction Modeling memory is not new; detailed descriptive models of memory have existed since the 1960s. Atkinson and Shiffrin (1968), Craik and Lockhart (1972), Tulving (1972), and, Baddeley and Hitch (1974), among others, are credited with providing the first early influential models of memory. A more recent phenomenon is the formalization of memory models using computers. Since memory does not exist in isolation from other components of the human mind, it is natural to embed memory within a larger scheme for modeling all of cognition. The concept of the cognitive architecture arose from this rationale. A cognitive architecture is a means for formalizing the principles that underlie one’s theory about how the mind works (Newell, 1990). It specifies the structure of the mind, including how information is gathered, processed, and used to perform tasks. There are several advantages to implementing a cognitive model. Using the architecture, models can be built which can then be used to simulate humans performing tasks. The resulting simulation data can be compared to human experimental data, which allows for both the evaluation of the particular model, and the forrnalizations specified by the architecture itself. Three of the most popular cognitive architectures are ACT—R (Anderson & Lebiere, 1998), SOAR (Laird, Newell & Rosenbloom, 1987; Newell, 1990), and EPIC (Kieras & Meyer, 1997). Of these, ACT—R has the best developed and well tested model of the human declarative memory system and is also highly consistent with mainstream theories about declarative memory from cognitive psychology. The declarative memory system of ACT—R (DM) arose out of a need to account for the results of experiments on human memory generated by experimental psychologists. DM is also a necessary component of accounts of most experiments on cognition, due to the need to account for participants’ memory of experimental instructions, and relevant general background knowledge. In developing an ACT—R model, a knowledge base relevant to the experimental task is usually provided to the model in advance of a simulation to represent knowledge held by the subject before beginning the experiment. During the experiment the model typically learns by creating new chunks and adjusting the relative activation of existing chunks (Anderson & Lebiere, 1998). This general scheme, and variations on it, has been very successful for modeling the sorts of tasks given to participants in psychological experiments related to memory and learning (Anderson & Lebiere, 1998). However, a unified theory of cognition ultimately needs to go beyond modeling individual tasks done in isolation (Newell, 1990). It must also account for how it is that we, as intelligent agents, accumulate a vast amount of world knowledge that we apply in a context specific manner to the various tasks that we perform throughout our lives. In order to make this distinction clear we will refer to the knowledge that we carry around outside of a particular context as persistent knowledge, and the specific knowledge needed for a particular task as task knowledge. Task knowledge is a local manifestation of persistent knowledge that has been retrieved to make it available for the task. ACT—R is a very good system for modeling how people use and manipulate task knowledge, but there is very little work focusing on the management of persistent knowledge. One approach that has been used for modeling how people process the enormous database that makes up persistent knowledge is the use of holographic memory models. The principle purpose of this paper is to present our Dynamically Structured Holographic Memory System (DSHM) as a means for modeling persistent knowledge and to suggest how it could be used to augment ACT—R. The format for the remainder of this paper is as follows: First, DSHM, based on the lexical representation system BEAGLE (Jones & Mewhort, 2007) is presented; second, a hybrid system, MALTA, which combines DSHM and ACT- 1433

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.