Dhumbal AI: Comparative Study of AI Agents

Overview

This repository implements AI agents for the Dhumbal card game, a South Asian multiplayer game with imperfect information. The project, detailed in "AI Agents for the Dhumbal Card Game: A Comparative Study", evaluates rule-based (Aggressive, Conservative, Balanced, Opportunistic), search-based (MCTS, ISMCTS), learning-based (DQN, PPO), and random agents through 1024-round tournaments. It includes game environment, agent implementations, statistical analysis, and visualizations, supporting game AI research and cultural preservation.

Features

• Game Environment: Python-based Dhumbal implementation (2–5 players).
• AI Agents: Random, rule-based, MCTS, ISMCTS, DQN, PPO.
• Evaluation: Win rate, economic performance, Jhyap success, decision efficiency.
• Analysis: Welch’s t-tests, Cohen’s d, 95% CI, visualized via Jupyter notebooks.
• Reproducibility: Fixed random seed (42), Python 3.9+, TensorFlow 2.8.

Installation

1. Clone Repository:

   git clone https://github.com/sahajrajmalla/dhumbal-ai.git
   cd dhumbal-ai

2. Virtual Environment:

   python -m venv venv
   source venv/bin/activate  # Windows: venv\Scripts\activate

3. Install Dependencies:

   pip install -r requirements.txt

Usage

• Quick Demo:

  python quick_demo.py

• Tournaments:

  • Rule-Based: python agents/rule_based/rule_based_agent.py
  • Search-Based: python agents/search_based/single_process_search.py
  • Learning-Based: python agents/learning_based/compete.py
  • Championship: python championship/final.py

• Visualizations:

  cd visualization
  jupyter notebook

  Run performance_metrics.ipynb and statistical_analysis.ipynb.

• Training:

  • DQN: python agents/learning_based/dqn/dqn.py
  • PPO: python agents/learning_based/ppo/ppo.py

Repository Structure

dhumbal-ai/
├── agents/
│   ├── rule_based/           # Rule-based agents and results
│   ├── search_based/         # MCTS, ISMCTS, and results
│   ├── learning_based/       # DQN, PPO, training, and results
│   └── hybrid/               # Placeholder for hybrid agents
├── championship/             # Cross-category tournament scripts and results
├── visualization/            # Notebooks and plots for analysis
├── quick_demo.py             # Demo script
├── README.md                 # This file
├── LICENSE                   # MIT License
└── .gitignore

Contributing

Fork, create a branch, commit changes, and submit a pull request. Follow PEP 8 and document code.

License

MIT License (see LICENSE).

Citation

Contact

Sahaj Raj Malla: [email protected]