Data Driven Metareasoning for Defending a Perimeter Against Cooperative Intrusion

Abstract Perimeter defense problem is a type of differential game, where a team of attacker agents try to score by intruding a perimeter region while a team of defenders try to defend the perimeter by intercepting the attackers. There are many approaches to the multi-agent perimeter defense problem proposed over the last decade with varying assumptions. However, even the most general approach cannot generate an optimal solution for every situation. In some situations, a heuristic-based solution might work better than a complex one. Choosing the right solution will help in higher performance and lower computation burden due to the simplicity of the solution. This paper discusses a metareasoning approach for the multi-agent perimeter defense problem. We introduce two data-driven algorithm selection policies, one based on exploratory data analysis and the other based on machine learning, that determine which algorithm to use given the initial state of the perimeter defense game. We generated a large set of perimeter defense instances, simulated different solution approaches, and used the results to derive metareasoning policies. The analysis of the solution approaches and the basis for metareasoning to determine the solution approach were based on both the performance of the approach and the computation time taken. The results from our simulation study show that using these metareasoning policies leads, on average, to better solutions than existing algorithms used separately.