In small scale multi-agent environments, every agent is aware of all of the others. This allows agents to evaluate the potential outcomes of their interaction for each of their possible interaction partners. However, this farsighted knowledge becomes an issue in large scale systems, leading to a combinatorial explosion in evaluation and is unrealistic in communication terms. Limited awareness of other agents is therefore the only plausible scenario in many large-scale environments. This limited awareness can be modeled as a sparse social network in which agents only interact with a limited subset of agents known to them. In this paper, we explore a model of dynamic multi-agent coalition formation in which agents are connected via fixed underlying social networks that exhibit different well known structures such as Small World, Random and Scale Free topologies. Agents follow different exploratory policies and are distributed in the network according to a variety of metrics. The primary results of the paper are to demonstrate different positive and negative properties of each topology for the coalition formation problem. In particular we show that despite positive properties for many problems, Small World topologies introduce blocking factors which hinder the emergence of good coalition solutions in many configurations.
展开▼
