Neural connect 4 - a connectionist approach to the game

摘要

This article presents the system "Neural Connect 4", a program that plays the game Connect Four. This system employs the multilayer perceptron architecture which is learning through the supervised backpropagation algorithm. The required knowledge for training comes from saved games. After a short introduction to the game itself the symbolic algorithms used for training and evaluation are described. Comparisons are made within a connectionist approach: effective and ineffective learning techniques are shown and the results are discussed. "Neural Connect 4" proves that artificial neural networks are completely adequate for learning Connect Four, given that certain principles are observed.