In this thesis, we study how reinforcement learning algorithms can tackle classical board games without recourse to human knowledge. Specifically, we develop a framework and algorithms which learn to play the board game Hex starting from random play. We first describe Expert Iteration (ExIt), a novel reinforcement learning framework which extends Modified Policy Iteration. ExIt explicitly decomposes the reinforcement learning problem into two parts: planning and generalisation. A planning algorithm explores possible move sequences starting from a particular position to find good strategies from that position, while a parametric function approximator is trained to predict those plans, generalising to states not yet seen. Subsequently, planning is improved by using the approximated policy to guide search, increasing the strength of new plans. This decomposition allows ExIt to combine the benefits of both planning methods and function approximation methods. We demonstrate the effectiveness of the ExIt paradigm by implementing ExIt with two different planning algorithms. First, we develop a version based on Monte Carlo Tree Search (MCTS), a search algorithm which has been successful both in specific games, such as Go, Hex and Havannah, and in general game playing competitions. We then develop a new planning algorithm, Policy Gradient Search (PGS), which uses a model-free reinforcement learning algorithm for online planning. Unlike MCTS, PGS does not require an explicit search tree. Instead PGS uses function approximation within a single search, allowing it to be applied to problems with larger branching factors. Both MCTS-ExIt and PGS-ExIt defeated MoHex 2.0 - the most recent Hex Olympiad winner to be open sourced - in 9 × 9 Hex. More importantly, whereas MoHex makes use of many Hex-specific improvements and knowledge, all our programs were trained tabula rasa using general reinforcement learning methods. This bodes well for ExIt’s applicability to both other games and real world decision making problems.
Expert iteration
2021-03-28
Doctoral thesis, UCL (University College London).
Theses
Electronic Resource
English
Manipulating the Distributions of Experience used for Self-Play Learning in Expert Iteration
BASE | 2020
|Wiley | 2009
|SPACE STATION: THE NEXT ITERATION
Online Contents | 1995
|Space Station: The Next Iteration
British Library Online Contents | 1995
|