Learning in Open Adaptive Networks

We propose a generic distributed learn-and-adapt model for high performance and high resilience configuration of open cooperative agent networks. Agents are involved into three interconnected types of activities. Firstly, agents bid for participation to the processing of a steady random flow of structured tasks. Secondly, agents learn the (exogenous) features of the random task source, by aggregating local information (such as success rates, average load, etc). And, thirdly, agents adapt the composition of their neighbourhoods following the (endogenous) targets set by their learning process. Neighbourhood readjustment proceeds by judicious rewiring steps which stay entirely local. Thus an agent continuously works, adjusts its neighbourhood, and based on his local metrics, learns how to inflect its own adaptation targets. Because of this tight coupling of all three activities, the network as a whole can reconfigure in a fully decentralized way to cope with changes in: the network composition (node failures, new incoming nodes, etc), and the parameters of the task source (changes in the size, structure, and frequency), while attaining robustly a near-optimal performance level (compared to the centralised solution).