Soar is a cognitive architecture developed at Carnegie Mellon University by John Laird, Allen Newell, and Paul Rosenbloom. It is being maintained and produced at the University of Michigan by John Laird's research group.

Overview

The Soar project seeks to create the fixed computational building blocks required for generally intelligent agents. Agents capable of performing a wide range of tasks and encoding, using and learning all types of knowledge to realise the full range of cognitive abilities in humans, such as decision-making, problem-solving, planning, and natural-language understanding. It is both a theoretical model of cognition and a computational application of that model. AI researchers have widely employed it to develop intelligent agents and cognitive models of various elements of human behaviour since its inception in 1983 as John Laird's thesis. The Soar Cognitive Architecture, published in 2012, is the most recent and thorough description of Soar.

Theory

Soar includes several theories regarding the computational structures that underpin general intelligence, many of which are shared with other cognitive architectures, such as ACT-R, developed by John R. Anderson, and LIDA, created by Stan Franklin. However, soar has recently focused on general AI (functionality and efficiency), whereas ACT-R has traditionally focused on cognitive modelling (detailed modelling of human cognition).

The Problem Space Hypothesis, described in Allen Newell's book Unified Theories of Cognition, is the basic cognition theory underlying Soar. We can trace it back to one of the original AI systems, Newell, Simon, and Shaw's Logic Theorist, which he initially showed in 1955.

Architecture

The connection between procedural memory (its knowledge of how to do things) and working memory (its representation of the current situation) in Soar's primary processing cycle supports the selection and application of operators. Working memory information is a symbolic graph structure rooted in a state. Procedural memory knowledge is represented as if-then rules (sets of conditions and actions) constantly matched against the contents of working memory. When a rule's conditions match structures in working memory, it triggers and executes its actions. A production system is another name for this mix of rules and working memory. However, unlike other production systems, all rules match the fire in parallel in Soar.

Conclusion

Soar's decision-making occurs through selecting and applying operators, which are proposed, assessed, and used by rules, rather than choosing a single rule. Rules that examine the current state and construct a working memory representation of the operator and an acceptable preference indicate that the operator should be considered for selection and application. Additional rules match with the proposed operator and generate preferences that compare and evaluate it to other suggested operators. 

A decision method analyses the preferences and selects the preferred operator, installed as the working memory's current operator. Rules that match the current operator are then executed to apply it and modify working memory. Simple inferences, queries for retrieval from Soar's long-term semantic or episodic memories, commands to the motor system to performing actions in an environment, or interactions with the Spatial Visual System (SVS), which is working memory's interface to perception, are all examples of changes to working memory. Working memory changes result in the proposal and evaluation of novel operators, followed by selecting one and its application.