A Distributed Asynchronous Constraint Optimization Algorithm Based on Dynamic Reduction of Constraint Graph

Distributed Constraint Optimization Problem (DCOP) is a framework that recently emerged as one of the most successful approaches for coordination in Multi-Agent system. Dynamic Programming Optimization (DPOP) is an algorithm which requires only a linear number of messages, thus introducing exponentially less network overhead than other search algorithms when applied in a distributed setting. In DPOP algorithm, each node should wait for UTIL messages from their child nodes. It causes only a small part of the agents are working while other agents have to stay in a waiting state. This restriction leads to a waste of computing resources. A new algorithm called Distributed Asynchronous Constraint Optimization (DACO) algorithm based on dynamic reduction of the constraint graph is proposed. The DACO algorithm allows nodes to send their partial solutions or constraints to another one and then be disappeared in a parallel way. In addition, it does not require transforming the constraint graph into a depth-first pseudo-tree (DFS tree) in advance as DPOP does. The algorithm together with the analysis is discussed in detail in the paper.