A Comparison of Approaches to Handling Complex Local Problems in DCOP

Many distributed constraint optimisation algorithms require each agent to have a single variable. For agents with multiple variables, there are two standard approaches: decomposition – for each variable in each local problem, create a unique agent to manage it; and compilation – compile the local problem down to a new variable whose domain is the set of all local solutions. We compare these two approaches with each other and with a modified compilation approach that uses dominance and interchangeabilities to reduce problem size and speed up search. Our preliminary results show: (i) the basic compilation is almost never competitive; (ii) the modified compilation gives significant improvements over the other methods as the size and complexity of each agent’s internal problem grows, as long as the number of inter-agent constraints and the domain size of the variables remains small; (iii) the decomposition approach is more appropriate to use as the number of inter-agent constraints and the domain size of the variables increase, as long as the overall problem size is small.