🧩 MCTS Predictor

A small machine learning project that predicts the outcome of Monte Carlo Tree Search (MCTS) variants when playing against each other in a given game.

The project:

• Trains a RandomForestRegressor on synthetic matchup data between pairs of MCTS parameter settings.
• Uses the trained model to predict win rates between variants.
• Runs round-robin tournament simulations and builds a simple Elo-style leaderboard of the variants.

📂 Project structure

mcts-predictor/
├── data/           # Generated datasets (CSV, created by train_model.py)
├── models/         # Saved ML models (.pkl, created by train_model.py)
├── train_model.py  # Generate synthetic data & train the model
├── tournament.py   # Simulate tournaments & build an Elo leaderboard
├── requirements.txt# Python dependencies
├── README.md       # Project documentation (this file)
└── .gitignore

Note: data/ and models/ are created automatically the first time you run train_model.py.

⚙️ Installation

Prerequisites

• Python: 3.9 or newer
• Recommended: a virtual environment (venv, conda, etc.)

Clone the repository and install dependencies

git clone <https://github.com/KomalGoel18/mcts-predictor>
cd mcts-predictor

python -m venv .venv

# Activate (Windows PowerShell)
.\.venv\Scripts\Activate.ps1

pip install -r requirements.txt

🏋️ Training the model

The training script:

• Generates a synthetic dataset of MCTS variant matchups.
• Trains a RandomForestRegressor to predict the win rate of Variant A vs Variant B.
• Saves the dataset to data/mcts_matchups.csv.
• Saves the trained model to models/mcts_model.pkl.

Run:

python train_model.py

Example console output (values will vary):

✅ Dataset saved to data\mcts_matchups.csv with 300 samples
📊 Model evaluation: MSE=0.0187, R²=0.6912
✅ Trained model saved to models\mcts_model.pkl

🎮 Running a tournament

Once the model has been trained and saved, you can run a simple round-robin tournament between predefined MCTS variants.

This will:

• Load the model from models/mcts_model.pkl.
• Define several variants (different values of exploration constant C and number of simulations).
• Predict win rates for every pair of variants.
• Build and print an Elo-style leaderboard.

Run:

python tournament.py

Example console output (truncated):

Matchup predictions:
    Player A   Player B  WinRate_A  WinRate_B
0  Variant A  Variant B   0.662917   0.337083
1  Variant A  Variant C   0.690833   0.309167
2  Variant A  Variant D   0.432083   0.567917
...
9  Variant D  Variant E   0.629583   0.370417

Leaderboard:
     Variant          Elo
0  Variant A  1513.689926
3  Variant D  1508.904551
1  Variant B  1503.485131
4  Variant E  1490.021374
2  Variant C  1483.899018

🧠 How it works (high level)

• Synthetic matches

  • train_model.py defines simulate_match, which takes two variants (each defined by C and number of simulations) and simulates a small number of games to estimate Variant A’s win rate.

• Features used for prediction

  • For each matchup, the dataset includes:
    • CA, simA: exploration constant and simulations for Variant A.
    • CB, simB: same for Variant B.
    • C_diff, sim_diff: differences between the two variants.
    • win_rate_A: empirical win rate of Variant A in the simulation.

• Model

  • A RandomForestRegressor is trained to predict win_rate_A from the above features.

• Tournament

  • tournament.py defines a list of named variants and uses the trained model to predict win rates for all pairs.
  • An Elo update rule is applied to build a leaderboard from these predicted outcomes.