Authors:
Dennis J. N. J. Soemers, Eric Piette, Matthew Stephenson, Cameron Browne

Venue:
Artificial Intelligence, 2023

Topics:
general game playing, spatial features, Monte Carlo Tree Search, explainable AI, pattern matching

Links: PDF · ScienceDirect · arXiv

This paper proposes the design and efficient implementation of spatial state-action features for general games.

These features describe local patterns around actions, and can be trained to incentivise or disincentivise moves based on the arrangement of pieces and empty positions in their neighbourhood.

The work emphasises both generality across many different board geometries and efficiency of evaluation, and the empirical study shows that these features can significantly improve the strength of agents using them to guide search.

This paper is a major contribution to the development of general game-playing agents that can learn and exploit local tactical patterns across many different games.

Rather than relying on game-specific handcrafted features or computationally expensive deep learning pipelines, the proposed approach aims for a balance between generality, efficiency, and interpretability.

It is particularly relevant for large-scale studies across many games in Ludii, and also for research directions related to explainable and human-like AI in games.

Soemers, D. J. N. J., Piette, E., Stephenson, M., Browne, C. (2023). Spatial State-Action Features for General Games. Artificial Intelligence.

@article{soemers2023spatial,
  author  = {Soemers, Dennis J. N. J. and Piette, Eric and Stephenson, Matthew and Browne, Cameron},
  title   = {Spatial State-Action Features for General Games},
  journal = {Artificial Intelligence},
  year    = {2023},
  url     = {https://arxiv.org/abs/2201.06401}
}