Quantifying the Expected Utility of Communicating Constraint Information * Avi Rosenfeld 1,2 , Sarit Kraus 2 and Charlie Ortiz 3 1 Department of Industrial Engineering Jerusalem College of Technology, Jerusalem, Israel 91160 2 Department of Computer Science Bar-Ilan University, Ramat-Gan, Israel 92500 3 SRI International, 333 Ravenswood Avenue Menlo Park, CA 94025-3493, USA Email: {rosenfa, sarit}@cs.biu.ac.il, ortiz@ai.sri.com November 30, 2007 Abstract In this paper we investigate methods for analyzing the expected value of communicating information in multi-agent planning and scheduling problems. As these problems are NP- complete, no polynomial algorithms exist for evaluating the impact a certain constraint, or relaxing the same constraint, will have on the global problem. We present a general approach where distributed agents estimate their problem tightness, or how constrained their local sub- problem is. Agents can then immediately identify many problems which are not constrained, and will not benefit from sending or receiving further information. Finally, agents use tradi- tional machine learning methods based on their specific local problem attributes to attempt to identify which of the constrained problems will most benefit from human attention. We evaluated this approach within a distributed c-TAEMS scheduling domain and found that this approach was overall effective. Keywords: Multiagent Scheduling, Adaptive Coordination, Localized Deci- sions * This research was supported by the DARPA Coordinators Program under Air Force Research Laboratory Con- tract FA8750-05-C-0033. Sarit Kraus is also affiliated with UMIACS.