Problem structure and subproblem sharing in multi-agent systems

In multi-agent problem-solving systems in which agents work asynchronously and in parallel on parts of a problem distributed among them, subproblem communication and integration protocols can have a major impact on problem solving efficiency. Transmitting too soon can necessitate having to "take back" transmitted partial solutions, wasting communication and processing time. Transmitting too late can complicate integration of partial solutions derived independently. The trade-off between these depends on the nature of the subproblem interdependency among the agents. We present a parameterized distributed constraint satisfaction problem (CSP) generator and a parameterized multi-agent problem-solving simulator: Using these tools, we offer some empirical observations of the tradeoff between strategies for subproblem communication timing. Our goal is to derive domain-independent answers to questions of subproblem sharing in multi-agent systems for problems that can be represented within the CSP paradigm.