Continuous Action Learning Automata Optimizer for training Artificial Neural Networks

Abstract

This paper introduces a Continuous-Action Learning Automata (CALA) game optimizer that provides a generalized way to use a game of CALA agents to train Artificial Neural Networks (ANNs) and Deep ANNs. This method uses both game theory and learning automata, which makes it a computationally efficient method when compared against other non-gradient and non-back propagation methods. Since the CALA game optimizer does not use gradients or back propagation, issues such as the vanishing gradient problem do not manifest, which allows for the use of multiple activation functions such as sigmoid or tanh even in a Deep ANN.