Towards a methodology for experiments with autonomous agents

Abstract

Experimental methodologies are harder to apply when selfmotivated agents are involved, especially when the issue of decision gains its due relevance in their model. Traditional experimentation has to give way to exploratory simulation, to bring insights into the design issues, not only of the agents, but of the experiment as well. The role of its designer cannot be ignored, at the risk of achieving only obvious, predictable conclusions. We propose to bring the designer into the experiment. We use the findings of extensive experimentation to compare current experimental methodologies in what concerns evaluation.1.