Causal Bayesian Optimization (CBO) - BoTorch Implementation

🎯 Project Overview

This project implements Causal Bayesian Optimization (CBO) algorithms using BoTorch, providing comprehensive comparison with standard Bayesian Optimization (BO) across multiple benchmark problems.

Project Origin: Adapted from VirgiAgl/CausalBayesianOptimization, migrated from GPy to BoTorch for better performance and modern PyTorch integration.

🚀 Quick Start

Environment Setup

Activate recommended environment

Required Dependencies:

• Python 3.11+
• PyTorch 2.5+
• BoTorch 0.10+
• GPyTorch 1.11+
• NumPy, Pandas, Matplotlib, Seaborn, Jupyter

Full Experiments

# Run complete CBO vs BO comparison (50 iterations, 5 seeds, 4 benchmarks)
python run_experiments.py

# View results
cat results/final_results.json

# Interactive analysis
jupyter notebook Result_Analysis.ipynb

🔧 Parameter Configuration

Main Scripts:

• BO_botorch.py: Standard Bayesian Optimization
• CBO_botorch.py: Causal Bayesian Optimization
• run_experiments.py: Automated batch experiments

Key Parameters:

• --graph_type: ToyGraph, CompleteGraph, CoralGraph, SimplifiedCoralGraph
• --num_trials: Optimization iterations (default: 50, recommended: 50-100)
• --seed: Random seed for reproducibility (0-9 recommended)
• --device: PyTorch device (cpu, cuda, auto)

Experiment Configuration (in run_experiments.py):

• Algorithms: ['BO', 'CBO']
• Seeds: [0, 1, 2, 3, 4] (5 replications)
• Iterations: 50 (production setting)
• Benchmarks: All 4 graph types