Self-adaptive Support Vector Machine: A multi-agent optimization perspective

Support Vector Machines (SVM) have been in the forefront of machine learning research for many years now. They have very nice theoretical properties and have proven to be efficient in many real life applications but the design of SVM training algorithms often gives rise to challenging optimization issues. We propose here to review the basics of Support Vector Machine learning from a multi-agent optimization perspective. Multi-agents systems break down complex optimization problems into elementary oracle tasks and perform a collaborative solving process resulting in a self-organized solution of the complex problems. We show how the SVM training problem can also be tackled from this point of view and provide several perspectives for binary classification, hyperparameters selection, multiclass learning as well as unsupervised learning. This conceptual work is illustrated through simple examples in order to convey the ideas and understand the behavior of agent cooperation. The proposed models provide simple formulations of complex learning tasks that are sometimes very difficult to solve with classical optimization strategies. The ideas that are discussed open up perspectives for the design of new distributed cooperative learning systems. (C) 2015 Elsevier Ltd. All rights reserved.