[Paper Review] Design and Implementation of TAG: A Tabletop Games Framework

Raluca D. Gaina, Martin Balla+3 more|arXiv (Cornell University)|Sep 25, 2020

Artificial Intelligence in Games6 references4 citations

TL;DR

This paper presents TAG, a Java-based framework for developing and benchmarking AI agents in modern tabletop games, offering a unified API, reusable game components, JSON-based game definition, and built-in analysis tools. It enables rapid prototyping of games like Tic-Tac-Toe, Love Letter, and Pandemic, and supports diverse AI algorithms including MCTS and RHEA, facilitating research in General Game AI, hidden information handling, and parameter tuning.

ABSTRACT

This document describes the design and implementation of the Tabletop Games framework (TAG), a Java-based benchmark for developing modern board games for AI research. TAG provides a common skeleton for implementing tabletop games based on a common API for AI agents, a set of components and classes to easily add new games and an import module for defining data in JSON format. At present, this platform includes the implementation of seven different tabletop games that can also be used as an example for further developments. Additionally, TAG also incorporates logging functionality that allows the user to perform a detailed analysis of the game, in terms of action space, branching factor, hidden information, and other measures of interest for Game AI research. The objective of this document is to serve as a central point where the framework can be described at length. TAG can be downloaded at: https://github.com/GAIGResearch/TabletopGames

Motivation & Objective

• Enable standardized research in General Game AI for modern tabletop games with diverse mechanics and hidden information.
• Provide a reusable, extensible framework that simplifies the implementation of new games and AI agents.
• Support both competitive and cooperative gameplay, including games with dynamic rule changes and asymmetric roles.
• Facilitate AI evaluation through built-in tournament systems, logging, and observability analysis (e.g., branching factor, action space).
• Encourage research in belief state learning, parameter tuning, and procedural content generation via extensible game and agent interfaces.

Proposed method

• Design a modular framework with core components (tokens, dice, cards, counters, boards) abstracted into reusable classes.
• Implement a game loop and forward model that supports state transitions via actions governed by rules and turn order.
• Provide a flexible API with abstract base classes for game state, actions, and rules, enabling easy game extension.
• Support game definition via JSON files for configuration and data import, reducing boilerplate in game development.
• Integrate logging and analysis tools to measure game complexity metrics such as branching factor, hidden information, and action space.
• Enable AI agent integration through standardized interfaces (e.g., IVectorObservation, IFeatureRepresentation) for reinforcement learning and planning algorithms.

Experimental results

Research questions

• RQ1 How can a unified framework support diverse tabletop games with varying mechanics, including hidden information and asymmetric roles?
• RQ2 To what extent can AI agents generalize across games with shared mechanics using transfer learning or parameter tuning?
• RQ3 How do different AI algorithms (e.g., MCTS, RHEA) perform across a range of competitive and cooperative tabletop games?
• RQ4 Can belief state construction be effectively modeled in AI agents when game states are partially observable?
• RQ5 How can procedural content generation and automated parameter tuning improve game balance and AI performance in complex tabletop games?

Key findings

• TAG successfully implements seven distinct tabletop games, including Love Letter, Uno, and Pandemic, demonstrating its versatility and extensibility.
• The framework enables detailed game analysis, including branching factor and hidden information metrics, supporting empirical evaluation of AI performance.
• AI agents using MCTS and RHEA achieve competitive performance across multiple games, with results available through built-in tournament and logging modules.
• The inclusion of vector and feature-based observations allows seamless integration with modern learning-based AI algorithms.
• The framework supports parameter tuning via the ITunableParameters interface, enabling systematic optimization of AI behavior and performance.
• The design encourages research into belief state learning, as AI agents must infer hidden information without historical state tracking, mirroring human cognition.