A primary challenge of agent-based reinforcement learning in complex and uncertain environments is escalating computational complexity with the number of the states. Hierarchical, or tree-based, state representation provides a promising approach to complexity reduction through clustering and sequencing of similar states. We introduce the Q-Tree algorithm to utilize the data history of state transition information to automatically construct such a representation and to obtain a series of linear separations between state clusters to facilitate learning. Empirical results for the canonical PuddleWorld problem are provided to validate the proposed algorithm; extensions of the PuddleWorld problem obtained by adding random noise dimensions are solved by the Q-Tree algorithm, while traditional tabular Q-learning cannot accommodate random state elements within the same number of learning trials. The results show that the Q-Tree algorithm can reject state dimensions that do not aid learning by analyzing weights of all linear classifiers for a hierarchical state representation. © Springer-Verlag Berlin Heidelberg 2012.
CITATION STYLE
Mao, T., Cheng, Z., & Ray, L. E. (2012). Q-Tree: Automatic construction of hierarchical state representation for reinforcement learning. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 7508 LNAI, pp. 562–576). https://doi.org/10.1007/978-3-642-33503-7_55
Mendeley helps you to discover research relevant for your work.