Relational Reinforcement Learning - Decreasing Game State Spaces Through Generalization

Jensen, Klaus

4. term

Computer Science, Master

2004

2012-02-14

Many machine learning problems are hard to solve due to the size of the state space used in the application. In such a case, finding the optimal solution requires a lot of computation. This report is part of a project, where focus lies on finding ways to decrease the size of state spaces used in small computer games. A commonly used machine learning technique known as Reinforcement Learning has a hard time dealing with large state spaces, because of the table-based Q-learning used to learn a given environment. Relational Interpretation can be used to extend conventional Reinforcement Learning with relational representation methods through First Order Logic, yielding the Relational Reinforcement Learning technique. The commonly used toy example of the Blocks World is used as example for showing the strenghts and weaknesses of Relational Reinforcement Learning. Finally, Tetris Limited, a reduced version of the well-known puzzle game of Tetris is presented and implemented using Reinforcement Learning.

Download
View record in AAU Student Projects