In-Time Agent-Based Vehicle Routing with a Stochastic Improvement Heuristic

Vehicle routing problems (VRP's) involve assigning a fleet of limited capacity service vehicles to service a set of customers. This paper describes an innovative, agent-based approach to solving a real-world vehicle-routing problem embedded in a highly dynamic, unpredictable domain. Most VRP research, and all commercial products for solving VRP's, make a static-world assumption, ignoring the dynamism in the real world. Our system is explicitly designed to address dynamism, and employs an in-time algorithm that quickly finds partial solutions to a problem, and improves these as time allows. Our fundamental innovation is a stochastic improvement mechanism that enables a distributed, agent-based system to achieve highquality solutions in the absence of a centralized dispatcher. This solution-improvement technology overcomes inherent weaknesses in the distributed problem-solving approach that make it difficult to find high-quality solutions to complex optimization problems. In previous work on similar problems, the MARS system of Fischer and Muller, et al., achieved an average route performance of roughly 124% of Solomon's algorithm for a VRP problem, which is known to achieve results that average roughly 107% of optimal. Our algorithm produces routes that average 106% those produced by an adaptation of Solomon's algorithm to a more general problem.