Self-organisation in an agent network via multiagent Q-learning

In this paper, a decentralised self-organisation mechanism in an agent network is proposed. The aim of this mechanism is to achieve efficient task allocation in the agent network via dynamically altering the structural relations among agents, i.e. changing the underlying network structure. The mechanism enables agents in the network to reason with whom to adapt relations and to learn how to adapt relations by using only local information. The local information is accumulated from agents' historical interactions with others. The proposed mechanism is evaluated through a comparison with a centralised allocation method and the K-Adapt method. Experimental results demonstrate the decent performance of the proposed mechanism in terms of several evaluation criteria.