Effect of Network Topology on Neighbourhood-Aided Collective Learning

Abstract

This article is about multi-agent collective learning in networks. An agent revises its current model when collecting a new observation inconsistent with it. While revising, the agent interacts with its neighbours in the community, and benefits from observations that other agents send on a utility basis. When considering the learning speed of an agent with respect to all the observations within the community, it clearly depends on the neighbourhood structure, i.e. on the network topology. A comprehensive experimental study characterizes this influence, showing the main factors that affect neighbourhood-aided collective learning. Two kinds of informations are propagated in the networks: hypotheses and counter-examples. This study also weights the impact of these propagation by considering some variants in which one kind of propagation is stopped. Our main purpose is to understand how network characteristics affect to what extent the agents learn and share models and observations, and consequently the learning speed within the community.