Agent-Collective Optimization through Influence Network Design

Tuesday, November 4, 2003 - 9:20am - 9:55am
Keller 3-180
Zoltan Toroczkai (Los Alamos National Laboratory)
The dynamics of human, and most biological populations is characterized by competition for resources. By its own nature, this dynamics creates the group of elites, formed by those agents who have strategies that are the most successful in the given situation, and therefore the rest of the agents will tend to follow, imitate, or interact with them, creating a social structure of leadership in the agent society. These inter-agent communications generate a complex social network with small-world character which itself forms the substrate for a second network, the action network. The latter is a dynamic, adaptive, directed network, defined by those inter-agent communication links on the substrate along which the passed information/prediction is acted upon by the other agents. By using the minority game for competition dynamics, here we show that when the substrate network is highly connected, the action network spontaneously develops hubs with a broad distribution of out-degrees, defining a robust leadership structure that is scale-free. Furthermore, in certain, realistic parameter ranges, facilitated by information passing on the action network, agents can spontaneously generate a high degree of cooperation making the collective almost maximally efficient.