Effective and Efficient Memory for Generally Intelligent Agents.

Abstract

Intelligent systems with access to large stores of experience, or memory, can draw upon and reason about this knowledge in a variety of situations, such as to improve the efficacy of their learning, decision-making, and actions in the world. However, little research has examined the computational challenges that arise when real-time agents require access to large stores of knowledge over long periods of time.

This dissertation explores the computational trade-offs involved in enhancing intelligent agents with effective and efficient memory. We exploit general properties of environments, tasks, and agent cues in order to develop scalable algorithms for episodic learning (autobiographical memory); semantic learning (context-independent store of facts and relations); and competence-preserving retention of learned knowledge (policies to forget memories while maintaining task performance). We evaluate these algorithms in Soar, a general cognitive architecture, for hours-to-days of real-time execution and demonstrate that agents with effective and efficient memory benefit along numerous dimensions when tasked within a variety of problem domains, including linguistics, planning, games, and mobile robotics.