Comparison of TDLeaf(lambda) and TD(lambda) Learning in Game Playing Domain

Daniel Osman, Jacek Mandziuk

Abstract

In this paper we compare the results of applying TD(λ) and TDLeaf(λ) algorithms to the game of give-away checkers. Experiments show comparable performance of both algorithms in general, although TDLeaf(λ) seems to be less vulnerable to weight over-fitting. Additional experiments were also performed in order to test three learning strategies used in self-play. The best performance was achieved when the weights were modified only after non-positive game outcomes, and also in the case when the training procedure was focused on stronger opponents. TD-learning results are also compared with a pseudo-evolutionary training method.
BibTeX
@inproceedings{osman2004comparison,
  author = {Daniel Osman and Jacek Mandziuk},
  title = {{Comparison of TDLeaf(lambda) and TD(lambda) Learning in Game Playing Domain}},
  booktitle = {ICONIP},
  year = {2004},
  doi = {10.1007/978-3-540-30499-9\_84},
}

Loading related content...

Figure