In deep reinforcement learning, searching and learning techniques are two important components. They can be used independently and in combination to deal with different problems in AI, and have achieved impressive results in game playing and robotics. These results have inspired research into artificial general intelligence (AGI), using these methods. Two general frameworks—General GamePlaying (GGP) and AlphaZero—have been built as the testbed to explore different aspects of AGI. Both frameworks combine searching and learning methods.
The purpose of this dissertation is to assess the potential of these methods. We study table based classic Q-learning on the GGP system, showing that classic Q-learning works on GGP, although convergence is slow, and it is computationally expensive to learn complex games.
This dissertation uses an AlphaZero-like self-play framework to explore AGI on small games. By tuning different hyper-parameters, the role, effects and contributions of searching and learning are studied. A further experiment shows that search techniques can contribute as experts to generate better training examples to speed up the start phase of training. This idea is called warm-start in the dissertation. We find that in AlphaZero-like self-play, a combination of Rollout and Rave enhancements can improve the startiterations of self-play training, especially with an adaptive iteration length.
In order to extend the AlphaZero-likeself-play approach to single player complex games, the Morpion Solitaire game is implemented by combining Ranked Reward method. Morpion Solitaire is a highly challenging combinatorial puzzle. Our first AlphaZero-based approach is able to achieve a near human best record. This result indicates that AlphaZero-like self-play approach is a promising method to explore AGI in single playergames.
Overall, in this thesis, both searching and learning techniques are studied (by themselves and in combination) in GGP and AlphaZero-like self-play systems. We do so for the purpose of making steps towards artificial general intelligence, towards systems that exhibit intelligent behavior in more than one domain. Our results are promising, and propose alternative ways in which search enhancements can be embedded as experts to generate better training examples for the start phase of training.
