[Paper Review] General Board Game Concepts

Éric Piette, Matthew Stephenson+2 more|arXiv (Cornell University)|Jul 2, 2021

Artificial Intelligence in Games65 references17 citations

TL;DR

This paper introduces a formal framework for 'game concepts'—high-level, human-understandable ideas like 'capture' or 'checkmate'—in General Game Playing (GGP) using the Ludii game system. By leveraging ludemes (atomic game components), the approach enables automatic detection of concepts across multiple abstraction levels, facilitating AI explainability, transfer learning, game generation, reconstruction, and recommendation.

ABSTRACT

Many games often share common ideas or aspects between them, such as their rules, controls, or playing area. However, in the context of General Game Playing (GGP) for board games, this area remains under-explored. We propose to formalise the notion of "game concept", inspired by terms generally used by game players and designers. Through the Ludii General Game System, we describe concepts for several levels of abstraction, such as the game itself, the moves played, or the states reached. This new GGP feature associated with the ludeme representation of games opens many new lines of research. The creation of a hyper-agent selector, the transfer of AI learning between games, or explaining AI techniques using game terms, can all be facilitated by the use of game concepts. Other applications which can benefit from game concepts are also discussed, such as the generation of plausible reconstructed rules for incomplete ancient games, or the implementation of a board game recommender system.

Motivation & Objective

• To formalize common game ideas—such as 'capture' or 'checkmate'—using terms familiar to players and designers.
• To enable automatic detection of these concepts across different levels of game abstraction (e.g., game, move, state) using Ludii’s ludeme-based game representation.
• To support new research in AI explainability, transfer learning, and game generation by providing a shared semantic layer across games.
• To address limitations in existing GGP systems like GDL, which lack high-level semantics and human interpretability.
• To enable practical applications such as game recommendation, rule reconstruction of ancient games, and benchmark selection for AI agents.

Proposed method

• Uses Ludii’s ludeme-based game description language to decompose games into atomic conceptual units.
• Defines game concepts as high-level, human-readable abstractions derived from ludemes, such as 'capture', 'check', or 'stalemate'.
• Automatically computes concepts at multiple abstraction levels: game-level (e.g., 'impartial game'), move-level (e.g., 'capture move'), and state-level (e.g., 'checkmate position').
• Leverages the structural and semantic consistency of ludemes to detect recurring concepts across diverse games.
• Applies concept detection to enable AI techniques such as move ordering, agent selection, and fitness evaluation in evolutionary game generation.
• Uses concept similarity across games to support rule reconstruction of incomplete historical games and to build game recommender systems.

Experimental results

Research questions

• RQ1 How can common game ideas such as 'capture' or 'checkmate' be formally defined and detected across diverse board games using a shared semantic framework?
• RQ2 To what extent can game concepts improve the explainability of AI techniques in General Game Playing?
• RQ3 Can game concepts enable effective transfer learning between games by identifying shared high-level features?
• RQ4 How can game concepts enhance game generation, reconstruction, and recommendation systems in GGP?
• RQ5 Can concept-based filtering improve the selection of benchmark games for evaluating AI agents?

Key findings

• Game concepts such as 'capture', 'check', and 'stalemate' can be automatically detected from ludeme-based game descriptions in Ludii, enabling consistent semantic analysis across games.
• The framework supports multi-level concept detection, including game-level (e.g., 'impartial game'), move-level (e.g., 'capture move'), and state-level (e.g., 'checkmate') concepts.
• Concept-based analysis enables the creation of a hyper-agent selector that chooses optimal AI agents based on game concepts, improving performance across diverse games.
• Game concepts facilitate transfer learning by identifying shared structural and strategic features between games, even when rules differ.
• The approach enables plausible rule reconstruction for incomplete ancient games by predicting missing concepts based on similar known games.
• Game concepts enhance recommender systems by allowing users to search for games based on preferred concepts (e.g., 'hidden information' or 'asymmetric play'), improving personalization and discovery.