Official implementation of "Online Robust Reinforcement Learning Through Monte-Carlo Planning" (ICML 2025).

This repository provides a robust variant of Monte Carlo Tree Search (MCTS) that addresses model ambiguity in both transition dynamics and reward distributions. Our algorithm bridges the gap between simulation-based planning and real-world deployment by incorporating distributionally robust optimization into the MCTS framework.

Key Features:

• Robust MCTS with multiple ambiguity sets (Total Variation, Chi-squared, Wasserstein)
• Non-asymptotic convergence guarantees matching standard MCTS (O(n^(-1/2)))
• Improved performance under model misspecification
• Implementation for Gambler's Problem environment

# Create conda environment
conda create -n rmcts python=3.9
conda activate rmcts

# Clone the repository
git clone https://github.com/brahimdriss/RobustMCTS.git
cd RobustMCTS

# Clone and install the rl-agents dependency
git clone https://github.com/brahimdriss/rl-agents.git
cd rl-agents
pip install -e .
cd ..

# Install required packages
pip install gym==0.26.2

.
├── agents/                  # MCTS agent implementations
│   ├── power_mcts/         # Standard Power-UCT agent
│   └── robust_mcts/        # Robust MCTS variants
├── environments/           # Environment implementations
│   └── gambler.py         # Gambler's Problem
├── experiments/           # Experiment scripts
│   └── run_comparaison.py # Main comparison experiments
├── utils/                 # Utility functions
│   ├── parallel.py       # Parallel execution utilities
│   └── visualization.py  # Plotting and visualization
└── tests/                # Unit tests

from environments.gambler import GamblerEnv
from agents.robust_mcts import RobustMCTSAgent

# Create environment
env = GamblerEnv(p_h=0.4, goal=10)

# Configure robust agent
config = {
    "c": 2,
    "gamma": 0.99,
    "power": 2,
    "budget": 2000,
    "max_depth": 50,
    "uncertainty_type": "wasserstein",  # or "tv", "chi2"
    "uncertainty_budget": 0.5
}

# Initialize agent
agent = RobustMCTSAgent(env, config)

# Get action
state, _ = env.reset()
action = agent.act(state)

The repository includes experiments on the Gambler's Problem, testing robustness to probability misspecification across different planning and execution environments. To run the comparison experiments:

python -m experiments.run_comparaison

Results are saved to the res/ directory with visualizations showing success rates under different levels of model mismatch.

@inproceedings{dam2025online,
title={Online Robust Reinforcement Learning Through Monte-Carlo Planning},
author={Tuan Quang Dam and Kishan Panaganti and Brahim Driss and Adam Wierman},
booktitle={Forty-second International Conference on Machine Learning},
year={2025}
}

• Authors: Tuan Quang Dam, Kishan Panaganti, Brahim Driss, Adam Wierman

For questions or issues, please open an issue on GitHub or contact the authors.