Genetric Programming to Learn an Agent''s Monitoring Strategy

Return-Path: weston@cs.umass.EDU Date: Wed, 31 Mar 93 09:10 EST From: Peggy Weston 31-Mar-1993 0910 Subject: Abstract 93-26 To: sutherla@freya.cs.umass.edu Message-Id: X-Envelope-To: sutherla@freya.cs.umass.edu X-Vms-To: sutherland@cs.umass.EDU Received: from freya.cs.umass.edu by cs.umass.EDU; Wed, 31 Mar 93 09:09 EST Received: by freya.cs.umass.edu (5.57/Ultrix3.0-C) id AA02541; Wed, 31 Mar 93 09:10:35 -0500 Date: Wed, 31 Mar 93 09:16:25 -0500 From: Peggy Weston Subject: Abstract 93-26 To: sutherland@cs.umass.EDU Message-id: Content-Type: TEXT/plain; charset=US-ASCII TR 93-26 -- Marc Atkin and Paul R. Cohen Many tasks require an agent to monitor its environment, but little is known about appropriate monitoring strategies to use in particular situations. Our approach is to learn good monitoring strategies with a genetic programming algorithm. To this end, we have developed a simple agent programming language in which we represent monitoring strategies as programs that control a simulated robot, and a simulator in which the programs can be evaluated. The effects of different environments and tasks is determined experimentally; changing features of the environment will change which strategies are learned. The correspondence can then be analyzed. We present the simulated robot with the task of getting as close to an obstacle as possible without touching it. We know of two strategies that might be used to solve this problem: periodic monitoring, and proportional reduction, in which the robot moves a fixed fraction of the remaining distance to the obstacle between monitoring events. Our hypotheses concerned the environmental conditions in which each strategy might be learned. An experiment confirmed some of our hypotheses and left the status of others ambiguous. At the same time, the experiment revealed new monitoring stragegies and set the stage for the further exploration of this field.