Coordination for uncertain outcomes using distributed neighbor exchange

Coordination of agent activites in non-deterministic, distributed environments is computationally difficult. Distributed Constraint Optimization (DCOP) provides a rich framework for modeling such multi-agent coordination problems, but existing representations, problem domains, and techniques for DCOP focus on small (<100 variables), deterministic solutions. We present a novel approach to DCOP for large-scale applications that contain uncertain outcomes. These types of real-time domains require distributed, scalable algorithms to meet difficult bounds on computation and communication time. To achieve this goal, we develop a new distributed neighbor exchange algorithm for DCOPs that scales to problems involving hundreds of variables and constraints and offers faster convergence to high quality solutions than existing DCOP algorithms. In addition, our complete solution includes new techniques for dynamic distributed constraint optimization and uncertainty in constraint processing. We validate our approach using test scenarios from the DARPA Coordinators program and show that our solution is very competitive with existing approaches.