Comparison of the Effectiveness of Simple Agent Capabilities for an On-line Area Coverage Task

We present the results of a suite of experiments conducted with a 2D simulation of a multi-agent area coverage problem. Agents perform a basic behavior of moving in a straight line and responding to collisions by setting a new random heading. Agents are optionally equipped with a subset of three additional capacities: a timer that causes the agent to alter its heading at regular intervals, a surface sensor that detects areas that have not yet been covered, and an agent sensor that detects the proximity of other agents. Our experimental conditions included all combinations of these features and also varied the task environment and the number of agents deployed. We found that the usefulness of feature configurations depends upon the task environment, number of agents deployed, and the metric used to determine performance. The surface sensor feature became more useful as the number of agents increased, while the agent proximity sensor became less useful. The surface sensor was generally useful early in the simulation, but became a hindrance once the task was nearly complete. With one performance metric, each of the three features alone was detrimental, while a configuration combining them together was beneficial.