Queuing Local Solutions in Distributed Constraint Satisfaction Systems

When solving Distributed Constraint Satisfaction Problems (DCSP), it is desirable that the search exploits asynchronism as much as possible so that the employed agents can perform much of the work in parallel to utilize the processing power available in a distributed environment. However, in many of todays DCSP algorithms only a few agents are working at any given time and the others are idling. This is caused by the fact that once an agent is consistent with its neighbors, it becomes idling until it is forced by other agents to choose a different assignment for its local variables. In this thesis we propose a method that utilizes the idling time of the agents to increase the efficiency of a distributed backtracking algorithm where agents have complex local problems and share variables among them. An agent computes solutions to its local problem in advance while it is waiting for incoming messages. This means that when an agent finds a solution to the local problem that is consistent with higher order agents, it not only informs lower order agents but continuous to search for further solutions which then are stored in a queue. When the current local solution becomes invalid due to a nogood received from a lower order agent, the agent does not have to search for a new local solution but can retrieve a precomputed one from the queue. This approach increases the amount of work the agents can perform in parallel since higher order agents search ahead for local solutions while lower order agents are trying to expand the current partial solution. Our experiments show that some increase in performance can be gained by queuing local solutions in distributed backtracking.