Abstract—This research paper presents an End-to-End Software Architecture based on Deep Neural Networks for Automatic Learning in Chess. Initially classifying and regressive approaches are explored to evaluating game configurations by employing deep belief networks. A third research approach which combines these, is then developed by quantizing the value range of the evaluation function. The neural network learns to assess game positions accurately during an unsupervised pre-training and supervised fine-tuning phase, using a dataset solely consisting of binary vector representations of the board and corresponding evaluations. An alpha beta tree search is used to complement the chess engine for finding optimal moves. The experiments show how artificial neural networks can develop a deep understanding of the application domain, despite having no prior knowledge of the game rules or strategies.
Index Terms—Chess, chess engine, artificial intelligence, deep learning, artificial neural 5networks, deep belief network, alpha beta pruning, alpha beta tree search.
K. Herud is with the Department of Applied Informatics, Baden-Wuerttemberg Cooperative State University, Lohrtalweg 10, 74821 Mosbach, Germany (e-mail: kon.herud.15@lehre.mosbach.dhbw.de).
C. Mueller is with the Department of Applied Informatics, Faculty of Informatics and Statistics of the University of Economics, W. Churchill Sq. 4, 130 67 Prague 3, Czech Republic (e-mail: research@ieoca.org).
[PDF]
Cite:Konstantin Herud and Carsten Mueller, "End-to-End Deep Neural Network for Automatic Learning in Chess," International Journal of Computer Theory and Engineering vol. 10, no. 5, pp. 146-151, 2018.
