Publication Type:Conference Paper
Source:ICCS 2011 - 8th International Conference on Complex Systems, NECSI Knowledge Press, Boston, MA, USA (2011)
In this paper, we propose an approach to reduce the number of interactions among agents in a multi-agent simulation. Modeling agent interactions turns out to be com- putationally expensive, especially when arbitrary interactions are allowed. In order to diminish computational costs of simulating agent interactions, we propose a self- organized approach to abstracting recurrent interaction patterns among groups of agents. Predictably interacting groups of agents are subsumed by higher-order agents that reproduce similar behaviours but at reduced computational costs. To this end, ob- server agents are immersed into the simulation space in order to monitor groups of agents and learn interaction patterns. Since the dynamics of the system changes over time, an abstraction might loose its validity and must therefore be removed again. This process is regulated by confidence values that are calculated and associated with individual abstractions. If a pattern exists for longer than anticipated, its confidence value is increased. The process of creating and removing abstractions is repeated dur- ing the course of a simulation in order to ensure an adequate adaptation to the system dynamics. Experimental results on a biological agent-based simulation show that our proposed abstraction method can successfully reduce the computational complexity during the simulation while maintaining the possibility of arbitrary interactions.