A probabilistic mechanism for agent discovery and pairing using domain-specific data

Agent discovery and pairing is a core process for many multi-agent applications and enables the coordination of agents in order to contribute to the achievement of organisational-level objectives. Previous studies in peer-to-peer and sensor networks have shown the efficiency of probabilistic algorithms in object or resource discovery. In this paper we maintain confidence in such mechanisms and extend the work for the purpose of agent discovery for useful pairs that eventually coordinate to enhance their collective performance. The key difference in our mechanism is the use of domain-specific data that allows the discovery of relevant, useful agents while maintaining reduced communication costs. Agents employ a Bayesian inference model to control an otherwise random search, such that at each step a decision procedure determines whether it is worth searching further. In this way it attempts to capture something akin to the human disposition to give up after trying a certain number of alternatives and take the best offer seen. We benchmark the approach against exhaustive search (to establish an upper bound on costs), random and tabu-all of which it outperforms-and against an independent industrial standard simulator-which it also outperforms. We demonstrate using synthetic data-for the purpose of exploring the resilience of the approaches to extreme workloads-and empirical data, the effectiveness of a system that can identify "good enough" solutions to satisfy holistic organisational service level objectives.