Introducing the Tileworld: Experimentally Evaluating Agent Architectures

Abstract

Abstract without regard to the amount of time it is taking or the changes meanwhile going on, is not likely to make ra.tional decisions. We describe a system called Tileworld, which con-sists of a simulated robot agent and a simulated environment which is both dynamic and unpre-dictable. Both the agent and the environment are highly parameterized, enabling one to control certain characteristics of each. We can thus ex-perimentally investigate the behavior of various meta-level reasoning strategies by tuning the pa-rameters of the agent, and can assess the success of alternative strategies in different environments by tuning the environmental parameters. Our hy-pothesis is that the appropriateness of a pa.rticu1a.r meta-level reasoning strategy will depend in large pa, rt upon the characteristics of the environment in which the agent incorporating that strategy is situated. We describe our initial experiments us-ing Tileworld, in which we have been evaluating a version of the meta-level reasoning strategy pro-posed in earlier work by one of the authors [Brat-man e2 al., 19SS]. One solution that has been proposed eliminates ex-plicit execution-time reasoning by compiling into the agent all decisions a.bout what to do in particular situations [Agre and Chapman, 1987, Brooks, 1987, Ka.elbling, 198S]. This is an interesting endeavor, but its ultimate feasibility for complex domains remains an open question.