Solution of Markov Reward Games Using Convolutional Neural Networks

Özkaya M., İzgi B.

12. International Conference on Applied Analysis and Mathematical Modeling, İstanbul, Turkey, 19 - 23 July 2024, pp.203-204

  • Publication Type: Conference Paper / Summary Text
  • City: İstanbul
  • Country: Turkey
  • Page Numbers: pp.203-204
  • Çanakkale Onsekiz Mart University Affiliated: Yes

Abstract

In this paper, we present a convolutional neural network architecture designed to solve Markov reward games. This architecture takes the rewards and transition matrix as inputs and provides the optimal strategy for the game. The proposed neural network architecture is trained using 80% of 3000 and 5000 Markov reward games, each featuring 3 actions and 3 states, and is tested utilizing 20% of 3000 and 5000 Markov reward games. The results reveal that the developed architecture can achieve errors of less than 3% in terms of mean square error in the final rewards.