Evolutionary Multi-Agent Systems: An Adaptive and Dynamic Approach to Optimization

This paper explores the ability of a virtual team of specialized strategic software agents to cooperate and evolve to adaptively search an optimization design space. Our goal is to demonstrate and understand how such dynamically evolving teams may search more effectively than any single agent or a priori set strategy. We present a core framework and methodology that has potential applications in layout, scheduling, manufacturing, and other engineering design areas. The communal agent team organizational structure employed allows cooperation of agents through the products of their work and creates an ever changing set of individual solutions for the agents to work on. In addition, the organizational structure allows the framework to be adaptive to changes in the design space that may occur during the optimization process. An evolutionary approach is used, but evolution occurs at the strategic rather than the solution level, where the strategies of agents in the team are the decisions for when and how to choose and alter a solution, and the agents evolve over time. As an application of this approach in a static domain, individual solutions are tours in the familiar combinatorial optimization problem of the traveling salesman. With a constantly changing set of these tours, the team, with each agent employing a different solution strategy, must evolve to apply the solution strategies, which are most useful given the solution set at any point in the process. We discuss the extensions to our preliminary work that will make our framework useful to the design and optimization community. [DOI: 10.1115/1.3013847]