Shruti-agent: a structured connectionist architecture for reasoning and decision-making

Abstract

One of the great scientific challenges of today is to understand how complex cognitive processes such as inference and decision-making can be achieved by the human brain. The incredible complexity of the brain makes computational modeling an indispensable tool to meet this challenge. Connectionist models that explain cognition in terms of brain-like structures and mechanisms are particularly useful in this endeavor. This thesis describes SHRUTI-agent, a neurally motivated, structured connectionist agent architecture that is capable of inference, decision-making, and interaction with a simulated environment. SHRUTI-agent is an extension of SHRUTI, a connectionist model that demonstrates how temporal synchrony variable binding in conjunction with structured, neurally plausible representations can support several aspects of high-level cognition. Creating a decision-making agent architecture based on SHRUTI involved a number of enhancements and additions to the SHRUTI model, and these comprise the primary contributions of the thesis. As a first step towards building an effective system for inference and decision-making, the evidential reasoning capability of the SHRUTI model was significantly enhanced. The resulting model demonstrates that relational knowledge and approximately probabilistic reasoning can be effectively combined in a connectionist architecture. Simple or reactive decision-making was made possible by the introduction into the model of compatible connectionist structures and mechanisms to represent and manipulate utility. It is shown that the basic decision-making capability of the model has much in common with formal methods from artificial intelligence. In order to enable the system to handle complex or sequential decision problems, a set of control mechanisms was developed. It is shown that the complex control mechanisms required for decision-making (e.g. hypothesis-testing) can be based on combinations of simple control primitives such as monitoring, filtering, and maintenance. Each of these control primitives is described in terms of its connectionist realization. Moreover, a subset of these control primitives are used as the basis for an implemented model of working memory in lateral prefrontal cortex. Operation of the SHRUTI-agent model is illustrated in two case studies, one involving simulated robot soccer and the other involving a military domain. Finally, consideration of the impact of learning on decision-making performance led to some preliminary work aimed at understanding how SHRUTI's structured representations can be learned. Utile concept learning is introduced as a mechanism for recruitment of new concepts and rules that lead to improved decision-making performance, and causal Hebbian learning is introduced as a mechanism for learning statistical causal rules.