A New Game Theory Based Approach for Self-Organizing Map Networks

Self-organizing map (SOM) is a well-known type of artificial neural networks (ANN), which is commonly used for vector quantization (VQ) and cluster analysis as well. Since the introduction of SOM, this method has been successfully applied to solve problems in various fields and many improvements and extensions are proposed. SOM uses a number of neurons to estimates the distribution of some input patterns in an n-dimensional space. Possible existence of dead neurons is a major problem of the SOM algorithm. Weight vectors of dead neurons are far from the input patterns, so they have no chance to compete with other neurons and contribute in the learning phase. Inappropriate initializations of neurons’ weights and non-convex shape of input distribution are the main causes of dead neurons. In this paper, the basic concepts of game theory are used and a new game theory based SOM algorithm is proposed in order to improve the map quality and solve the dead neuron problem. Each neuron is considered as a player with a set of strategies. During the learning phase, players compete with each other to obtain more input patterns. The proposed algorithm is then applied to some benchmark data distributions. The simulation results easily approve the effectiveness of proposed approach.

Keywords:

Self-organizing Map (SOM) , Neural Gas (NG) , Game Theory , Dead Neurons