Ensemble case based learning for multi-agent systems

This monograph presents a framework for learning in a distributed data scenario with decentralized decision making. We have based our framework in MultiAgent Systems (MAS) in order to have decentralized decision making, and in Case-Based Reasoning (CBR), since the lazy learning nature of CBR is suitable for dynamic multi-agent systems. Moreover, we are interested in autonomous agents that collaboratively work as ensembles. An ensemble of agents solves problems in the following way: each individual agent solves the problem at hand individually and makes its individual prediction, then all those predictions are aggregated to form a global prediction. Therefore, in this work we are interested in developing ensemble case based learning strategies for multi-agent systems. Specifically, we will present the Multi-Agent Case Based Reasoning (MAC) framework, a multi-agent approach to CBR. Each individual agent in a MAC system is capable of individually learn and solve problems using CBR with an individual case base. Moreover, each case base is owned and managed by an individual agent, and any information is disclosed or shared only if the agent decides so. Thus, this framework preserves the privacy of data, and the autonomy to disclose data. The focus of this thesis is to develop strategies so that individual learning agents improve their performance both individually and as an ensemble. Moreover, decisions in the MAC framework are made in a decentralized way since each individual agent has decision autonomy. Therefore, techniques developed in this framework achieve an improvement of individual and ensemble performance as a result of individual decisions made in a decentralized way. Specifically, we will present three kind of strategies: strategies to form ensembles of agents, strategies to perform case retention in multi-agent systems, and strategies to perform case redistribution.