dueling-bandit 0.1.0

A toolkit for preference-based online learning with dueling bandits

Dueling Bandit Toolkit

A Python package for preference-based online learning with dueling bandits. This toolkit implements algorithms like Double Thompson Sampling, PARWiS, and Contextual PARWiS, with support for real-world datasets (Jester, MovieLens) and evaluation metrics. Features

Algorithms: Double Thompson Sampling, PARWiS, Contextual PARWiS, and Random Pair baseline. Environment: Bradley-Terry model with optional contextual features. Datasets: Synthetic, Jester, and MovieLens support. Metrics: Cumulative regret, recovery fraction, true/reported ranks, and separation (Δ_1,2). Visualization: Plotting functions for experiment results.

Installation pip install dueling-bandit

Usage from dueling_bandit.environment import DuelingBanditEnv from dueling_bandit.agents import DoubleThompsonSamplingAgentpython -m build from dueling_bandit.experiments import run_simulation

Create environment

env = DuelingBanditEnv.random_bt(k=20, d=5, seed=42)

Initialize agent

agent = DoubleThompsonSamplingAgent(k=20, seed=42)

Run simulation

results = run_simulation(env, agent, horizon=500)

Plot results (requires matplotlib)

from dueling_bandit.plotting import plot_metric plot_metric({'500': {'Double TS': results}}, budget=500, dataset='synthetic', metric='mean_regret')

Requirements

Python >= 3.8 numpy, matplotlib, scipy, pandas

Install dependencies: pip install -r requirements.txt

Development

Clone the repository:git clone https://github.com/shailendrabhandari/dueling_bandit.git cd dueling-bandit

Install in editable mode:pip install -e .

Run tests:pytest tests/

Documentation Full documentation is available at ReadTheDocs. License MIT License. See LICENSE for details. Contributing Contributions are welcome! Please open an issue or pull request on GitHub.