Measuring the expected gain of communicating constraint information

Abstract

In this paper we investigate methods for measuring the expected utility from communicating information in multi-agent planning and scheduling problems. We consider an environment where human teammates can potentially add information to relax constraint information. As these problems are NP-complete, no polynomial algorithms exist for evaluating the impact of either adding or relaxing a certain constraint will have on the global problem. We present a general approach based on a notion we introduce called problem tightness. Distributed agents use this notion to identify those problems which are not overly constrained and, therefore, will not benefit from additional information that would relax those constraints. Finally, agents apply traditional machine learning methods based on their specific local problem attributes to attempt to identify which of the constrained problems will most benefit from added information. We evaluated this approach within a distributed c-TAEMS scheduling domain and found that this approach was effective overall.