Fully connected multi-objective particle swarm optimizer based on neural network

Abstract

In this paper, a new model for multi-objective particle swarm optimization (MOPSO) is proposed. In this model, each particle's behavior is influenced by the best experience among its neighbors, its own best experience and all its components. The influence among different components of particles is implemented by the on-line training of a multi-input Multi-output back propagation (BP) neural network. The inputs and outputs of the BP neural network are the particle position and its the 'gradient descent' direction vector to the less objective value according to the definition of no-domination, respectively. Therefore, the new structured MOPSO model is called a fully connected multi-objective particle swarm optimizer (FCMOPSO). Simulation results and comparisons with exiting MOPSOs demonstrate that the proposed FCMOPSO is more stable and can improve the optimization performance. © 2011 Springer-Verlag.