PROJECT TITLE :

Online Adaptable Learning Rates for the Game Connect-4

ABSTRACT:

Learning board games by self-play encompasses a long tradition in computational intelligence for games. Based mostly on Tesauro's seminal success with TD-Gammon in 1994, several successful agents use temporal distinction learning nowadays. However so as to achieve success with temporal difference learning on game tasks, often a careful selection of options and a massive number of coaching games is important. Even for board games of moderate complexity like Connect-4, we found in previous work that a very made initial feature set and several numerous game plays are needed. In this work we tend to investigate completely different approaches of online-adaptable learning rates like Incremental Delta Bar Delta (IDBD) or temporal coherence learning (TCL) whether or not they need the potential to speed up learning for such a advanced task. We propose a new variant of TCL with geometric step size changes. We tend to compare those algorithms with several different state-of-the-art learning rate adaptation algorithms and perform a case study on the sensitivity with respect to their meta parameters. We show that during this set of learning algorithms those with geometric step size changes outperform those different algorithms with constant step size changes. Algorithms with nonlinear output functions are slightly better than linear ones. Algorithms with geometric step size changes learn faster by a factor of four as compared to previously printed results on the task Connect-4.