Training learning agents to play computer games

OData support
Engedy István Tamás
Department of Measurement and Information Systems

In my thesis, I study the use of reinforcement learning in a computer game. For this expirement, I use Tetris as an environment, as this is a one player game, in which the agent can operate by observing only the environment, wtihout having the difficulty of other agents' present. The task of the agent is, to find an optimal policy in the chosen game.

A state description is required to the game, which will be the base of the reinforcement learning. In addition to this, the agent has to take rewards for its actions. I will review the weighting of the rewards and other parameters of the learning, to see how can the agent learn the game with different settings.

I evaluate the learning speed os the agent, depending on the parameters. Then I will review the quality of the learned policy, based on the points the agent earned during a game. The algorithm realated to reinforcement learning will be tested be specific versions of the game. Finally, I use the results to evaluate the chosen state description.


Please sign in to download the files of this thesis.