Original title: Fast and Knowledge-Free Deep Learning for General Game Playing (Student Abstract)
Authors: Michał Maras, Michał Kępa, Jakub Kowalski, Marek Szykuła
In this article, the authors describe a new approach to adapting the AlphaZero model for General Game Playing (GGP). They aim to make the model generation process faster and require less knowledge to be extracted from the game rules.
Instead of using self-play for dataset generation, the authors use a technique called Monte Carlo Tree Search (MCTS) playing. Additionally, they only use the value network and replace the convolutional layers with attention layers. This allows them to eliminate assumptions about the action space and board topology.
The authors implemented their method within the Regular Boardgames GGP system and conducted tests to evaluate the performance of the models they built. The results showed that their models outperformed the UCT baseline in most games and were able to achieve these results efficiently.
Overall, this article introduces a novel approach to adapting the AlphaZero model for GGP, showcasing its potential to improve model generation and performance in various games.
Original article: https://arxiv.org/abs/2312.14121