Prediction and Planning: Bridging the Gap

Duke University, Durham NC

Investigators

Abstract

This project is fundamental research to improve the performance of intelligent software agents, based on the observation that an agent's past experiences are a valuable and generally underutilized database. The goal is to produce algorithms that make stronger use of data than existing reinforcement learning algorithms, enabling a view of the agent's stored experiences as a repository that can be mined for performance-improving information. More generally, the agent may choose to use data obtained by observing other agents, or even from mining the web. The impact of this research may be felt in many areas. For example, software learning agents can be expected to learn in a much more human-like manner; noteworthy experiences will be remembered, and their influence on future performance will not attenuate. There will be no sampling requirements on the data, so it will be possible to learn from watching others and possible to use repositories of stored data to learn new behaviors. Among the likely practical applications of this work are network management and electronic commerce.