A comparison of two sigmoidal-type activation functions in video game controller evolution

Abstract

This paper presents an empirical comparison of two sigmoidal-type activation functions in evolutionary artificial neural network models. They are the log-sigmoid and hyperbolic tangent sigmoid activation functions which were investigated in order for evolving neural network controllers to play a classic video game. A Hill-Climbing Neural Network (HillClimbNet) was developed using the hill-climbing method together with a feedforward neural network to automatically create an intelligent controller that can play the screen-capture of Ms. Pac-man arcade game. The experimental results showed that that the HillClimbNet with log-sigmoid outperforms the HillClimbNet with hyperbolic tangent sigmoid when used in the hidden and output layers of the network when the agent plays the game.