Simulating UAV Surveillance for Analyzing Impact of Commitments in Multi-Agent Systems

Introduction

Rational agents execute actions that are in their “best interests” according to their desires. Rational action selection, already a difficult problem (in both representation and reasoning), is complicated by dynamic environments. This includes, in the case of multi-agent systems, dynamism resulting from interactions with other agents. Dynamism introduces uncertainty about the effects of actions, obfuscating the true value of each action. In the case of dynamism introduced through agent interactions, this may be somewhat mitigated if an agent can to predict future effects based on knowledge about the agent interactions that have occurred. In ad hoc agent systems, where the agents were not designed to plan together, this knowledge is gained through the announcement of agent intentions, resulting in social commitments among the agents. Commitments direct and constrain an agent’s future actions.

There has been much work on studying the semantics of commitment (Jennings, 1993). The objective of this research is to analyze the pragmatics of commitment in agent action and multi-agent coordination for the purposes of furthering our understanding of how to design and program coordinating agents. The approach taken is to explicitly model and reason about agent desires, specifically calculating the change in expected utility due to commitments. Commitments may have a negative impact on an agent’s ability to satisfy its desires. An accurate valuation of commitment is needed to determine if it is rational for an agent to commit to a goal or not. The domain of Unmanned Aerial Vehicle (UAV) surveillance is used as the representative domain for a class of problems addressed in this research. Simulation of this UAV domain is used to experimentally test this research.