Markantonis Stavros (Phd Candidate)

Thesis title: Reinforcement learning algorithms and mixture of experts architectures for efficiently solving stochastic problems with large state spaces
Supervisor: Diamantaras Konstantinos
Advisory Committee Members:
Goulianas Konstantinos, Professor, Dept. of Information and Electronic Engineering, IHU
Tefas Anastasios, Professor, School of Informatics, AUTH

The goal of the research that will be conducted is the implementation of Reinforcement Learning algorithms that will effectively explore the state space of complex stochastic problems by attributing to these states a state value close to the real one. The developed algorithms will not exploit hand-crafted special features, will not apply proactive searches, and will not require high computational training costs. Searching for the optimal model architecture, the results of these algorithms will be used to train various Mixture of Experts solutions. To test the results of the generated algorithms and architectures, it is proposed to apply them to stochastic games with large state spaces. The use of backgammon is considered, which is particularly recommended for testing pattern recognition methods in noisy, stochastic conditions. A draw is impossible, the game will eventually end even with random play, and the stochasticity of the dice will drive state exploration more thoroughly than in a deterministic game. The generated models can be tested against Pubeval, against GNU Backgammon with opponents of different levels, and at the International Computer Games Association’s Computer Olympiad.