Supervised and Reinforcement Learning in Neural Network Based Approach to the Battleship Game Strategy

Abstract

In our study the Battleship game we concern as an example of a simple pattern matching problem in correspondence with the Partially observable Markov decision process. We provide comparison of supervised and reinforcement learning paradigms used as neural network learning mechanisms applied by solving the Battleship game.We examine convergence of the neural network adaptation process by using these techniques.While concerning our pattern matching problem of the Battleship game solution by the neural network the reinforcement learning technique is not as straightforward as the supervised learning. On the other hand the neural network adaptation by the supervised learning mechanism has a faster convergence in our case. We use the Battleship game probability model to determine next position in an environment to be shot at with the highest probability of resulting into a successful hit attempt. © Springer International Publishing Switzerland 2013.