Authors:
Cameron Browne, Dennis J. N. J. Soemers, Éric Piette

Venue:
AAAI Workshop on Knowledge Extraction from Games (KEG), 2019

Topics:
General Game Playing, MCTS, feature learning, game AI, explainable AI

Links: PDF · CEUR Workshop Proceedings

This paper presents an approach for automatically learning strategic features for general games in the context of the Digital Ludeme Project.

The method focuses on defining geometric patterns that can bias Monte Carlo Tree Search (MCTS) playouts across a wide variety of games, independently of their underlying board geometry.

These features enable knowledge transfer between games and provide a basis for extracting human-interpretable strategies from learned behaviours.

This work is part of the Digital Ludeme Project and contributes to the development of the Ludii general game system.

It addresses a key challenge in General Game Playing: how to incorporate domain knowledge in a general and transferable way across different games.

The proposed feature-based approach bridges the gap between purely random simulations and domain-specific heuristics, while also opening perspectives for explainable AI in game-playing systems.

Browne, C., Soemers, D. J. N. J., Piette, É. (2019). Strategic Features for General Games. AAAI Workshop on Knowledge Extraction from Games (KEG), CEUR Workshop Proceedings, Vol. 2313.

@inproceedings{browne2019strategic,
  author    = {Browne, Cameron and Soemers, Dennis J. N. J. and Piette, Éric},
  title     = {Strategic Features for General Games},
  booktitle = {Proceedings of the AAAI Workshop on Knowledge Extraction from Games (KEG)},
  series    = {CEUR Workshop Proceedings},
  volume    = {2313},
  year      = {2019},
  url       = {https://ceur-ws.org/Vol-2313/}
}