TLAB, task learning architecture using behaviors, and its application to a mobile agent

Andrew Haydn Jones, Purdue University

Abstract

It is now commonly believed that intelligent agents are best implemented via multiple behaviors. For example, in the context of mobile agents, the various behaviors may correspond to “follow hallway”, “avoid obstacles”, “seek goal”, etc. It is not difficult to see that in any tangible application, the various behaviors will produce a discordant set of commands for the robot. The problem then becomes one of arbitrating between the commands dictated by the various behaviors and choosing one of the commands or some combination of them. This problem is known as the problem of arbitration. In the past, various solutions have been suggested for solving this problem such as the vector field approach by Arkin [1], the subsumption approach by Brooks [2], and the task based approach by Simmons [3]. In our research, we present a different solution. We think of behaviors as mapping functions from the domain that consists of sensory readings to the co-domain that consists of multiple sets of command decisions. In our approach, the behaviors are combined by taking a weighted sum of these functions. The individual weights are initially specified, but dynamically modified and learned by a reinforcement learning scheme. All these behaviors are programmed within a framework which is another aspect of our research. The results, both from simulation and real world situations, support our claim of solving the problem of arbitration while providing the system with the ability to dynamically adapt to hitherto unknown environments.

Degree

Ph.D.

Advisors

Kak, Purdue University.

Subject Area

Electrical engineering

COinS