Distributed Constraint Programming with Agents

Many combinatorial optimization problems lend themselves to be modeled as distributed constraint optimization problems (DisCOP). Problems such as job shop scheduling have an intuitive matching between agents and machines. In distributed constraint problems, agents control variables and are connected via constraints. We have equipped these agents with a full constraint solver. This makes it possible to use global constraint and advanced search schemes. By empowering the agents with their own solver, we overcome the low performance that often haunts distributed constraint satisfaction problems (DisCSP). By using global constraints, we achieve far greater pruning than traditional DisCSP models. Hence, we dramatically reduce communication between agents. Our experiments show that both global constraints and advanced search schemes are necessary to optimize job shop schedules using DisCSP.