|
|
|
|
|
ABSTRACT In recent years there has been growing interest in using game-dynamical mechanisms for controlling large-scale multi-agent systems (MAS). Such systems may possess certain global properties that can not be simply deduced from details of the microscopic characteristics of individual agents, but arise out of interactions among many agents. In particular, it has been established that in a system of competitive agents with bounded rationality, simple interaction rules can lead to a globally optimal (or near-optimal) system performance, a phenomenon commonly termed as Emergent Coordination. In this talk I will present a model of game-dynamical MAS that demonstrates emergent coordination and is, moreover, very robust and adaptive to changes in the environment. Our model consists of locally interacting boolean agents that compete for a limited resource. The capacity of the resource is allowed to vary externally, mimicking a dynamic environment. At each time step the agents face a binary choice of whether to use the resource or not, and those who use the resource are rewarded (punished) if their number is less (greater) than resource capacity. The framework of the inter-agent interactions is based on random boolean networks (Kauffman's NK model), where each agent (node) gets its input from K other randomly chosen agents, and maps the input to a new state according to a boolean function of K variables. The generalization we make is that the agents are allowed to adapt by having more than one boolean function, or strategy, and the use of a particular strategy is determined according to a simple reinforcement learning scheme. Our results indicate that for some parameters of the network the system shows a tendency towards self organization into a phase characterized by very effective utilization of the resource, even for relatively large variations in the capacity level. Remarkably, this parameter range corresponds to the boundary of ordered/chaotic phases in the prototype boolean networks. I will also talk about a generalization of our model for a multiple-resource case, and consider other potential applications of adaptive boolean networks, such as solving kSAT problems. SPEAKER BIO Aram Galstyan is a research associate at the USC's Information Sciences Institute. He received his Ph.D. in theoretical physics from the University of Utah, in 2000. His current research is focused on the use of statistical physics and numerical methods for analyzing emergent phenomenon in large scale multi-agent systems. Aram Galstyan |
|
Back to CRES |