Commitment and Eeectiveness of Situated Agents

Recent research in real-time Arti cial Intelligence has focussed upon the design of situated agents and, in particular, how to achieve e ective and robust behaviour with limited computational resources. A range of architectures and design principles has been proposed to solve this problem. This has led to the development of simulated worlds that can serve as testbeds in which the e ectiveness of di erent agents can be evaluated. We report here an experimental program that aimed to investigate how commitment to goals contributes to e ective behaviour and to compare the properties of di erent strategies for reacting to change. Our results demonstrate the feasibility of developing systems for empirical measurement of agent performance that are stable, sensitive, and capable of revealing the e ect of \high-level" agent characteristics such as commitment. Such systems are likely to have an increasing role to play in guiding the design of situated agents for speci c domains, and in contributing to a better understanding of how the characteristics of agents and environments interact.