Context-Aware Inefficiency of Self-Interested Agents in Unmanned Aerial Vehicle Wildfire Fighting

Multi-agent systems are present in many real-world aerospace applications, such as firefighting with unmanned aerial vehicles, making the development of autonomous agents that can solve complex tasks an active area of research. We focus on systems where a set of self-interested agents interact with each other and their environment in a manner that affects a desired social outcome, such as minimizing the number of burned trees in the firefighting example. Given such systems, our goal is to measure and understand how these behaviors might induce inefficiencies in the social outcome—we specifically aim to do so in a manner that considers the context within which these inefficiencies occur. Understanding this context could, e.g., help a system operator intervene when the system approaches a situation that is predicted to have high inefficiency or help a system designer ensure such situations are unlikely to occur. We address this problem by introducing a state-dependent price of anarchy (sPoA) metric that can be used within a sequential decision-making problem formalized as a Markov game. We then outline a computational approach for estimating this metric and understanding observed trends from empirical experiments. We demonstrate our approach in a multi-agent unmanned aerial vehicle firefighting task. Our empirical findings show that sPoA identifies nontrivial trends in inefficiencies due to self-interested behaviors across different states, highlighting the importance of considering state dependence in this problem.