Offline Multi-Objective Bandits

This repository contains the official implementation of our AAAI 2026 paper: Offline Multi-Objective Bandits: From Logged Data to Pareto-Optimal Policies

📁 Contents

• synthetic_main.py — main experiment runner.
• data/ — synthetic datasets and generator (data/synthetic_data.py).
• algorithms/ — algorithm implementations (OffMOB, MOLinLCB, MOKernLCB, neural baselines, ...).
• core/ — experiment utilities, runner, metrics, plotting.
• results/ — saved .npz experiment outputs (created by the runner).

🧰 Requirements

• Python 3.8+
• Recommended packages: numpy, scipy, torch, tqdm, easydict, absl-py, matplotlib

▶️ Quick start

Run a small OffMOB experiment:

python synthetic_main.py --algo_group=offmob --num_steps=100 --num_sim=2 --data_type=quadratic --results_dir=results

Important flags (see top of synthetic_main.py):

• --algo_group : all / offmob / baseline / ablation
• --data_type : quadratic / quadratic2 / cosine
• --noise_variance, --num_steps, --num_sim, --num_actions, --context_dim, --num_objectives
• --results_dir : output folder for .npz files

📊 Results format

Runners save .npz files (e.g. *_final.npz) containing:

• results : dict mapping algorithm name → metrics arrays (tchebycheff gaps, accuracies, rewards)
• test_points : list of evaluation steps

🛠️ Development notes

• Seeds and hyperparams live in the runner; create_hparams configures defaults.
• To add an algorithm implement sample_action, update, reset in the algorithm class.
• Save backups before overwriting large .npz result files.

📚 Cite

If you use this code, please cite the paper: