Files
Abstract
Individual planning in open multiagent systems, which involve a large number of agents and no communication medium among them, is a particularly difficult problem to solve due to severe uncertainty and exponential computational complexity with each added agent. Furthermore, the agent openness requires an agent to predict the presence of other agents as each agent can leave or rejoin the environment during the operation, and wrongly predicting an agents presence can add to a non-optimal behavior. For example, autonomous firefighting robots having no communication medium among them, tasked with fighting wildfires, may run out of suppressants and be temporarily unavailable to assist their peers. That requires agents to predict not only the actions of all the firefighting agents yet their suppressants levels as well, which becomes computationally intractable with the increase in the number of agents. To solve such a complex problem, this research proposes a novel method in this context that enables an agent to scalably and individually reason about others presence as well as their actions in its shared environment. With effective use of well-established sampling approaches in statistics, the method uses a principled approach to sample only a subset of neighboring agents and extrapolate their models to the overall population and combines it with an extension of Monte Carlo tree search to individual agent reasoning in multiagent environments. Simulations of multiagent wildfire suppression experiments demonstrate the potency of the new framework compared with alternative baseline methods and manifest the new framework as a reliable methodology for individual planning in multiagent systems with a large number of agents. The results exhibit the utility of the sampling methods in reducing the effects of the agents facing the Volunteers dilemma and directionless mental models.