Multi-Objective Causal Bayesian Optimization

Authors: Shriya Bhatija, Paul-David Zuercher, Jakob Thumm, Thomas Bohné

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic and real-world causal graphs demonstrate the superiority of our approach over non-causal multi-objective Bayesian optimization in settings where causal information is available.
Researcher Affiliation Academia 1Department of Computer Engineering, Technical University of Munich, Munich, Germany 2Department of Engineering, University of Cambridge, Cambridge, United Kingdom 3The Alan Turing Institute, London, United Kingdom. Correspondence to: Shriya Bhatija <EMAIL>.
Pseudocode Yes We propose our algorithm to solve MO-CBO problems1, for which the procedure is summarized in Algorithm 1. It assumes a known causal graph G, Y, X, C , prior data D, and a set S {OG,Y, MG,Y, P(X)} that specifies which local problems to consider.
Open Source Code Yes 1The full implementation of our algorithm is available at https://github.com/Shriya Bhatija/MO-CBO
Open Datasets Yes The model is inspired by the German Credit UCI dataset (Murphy, 1994), with causal dependencies adapted from Karimi et al. (2020). ... Murphy, P. M. UCI repository of machine learning databases, 1994. URL ftp://ftp.ics.uci.edu/ pub/machine-learning-databases/. ... This model originates from previous works of Ferro et al. (2015), and is based on real-world causal relationships.
Dataset Splits No We assume to have an initial dataset D = {((Xs, xk s), µ(Xs, xk s))}K,|S| k=1,s=1 with K = 5 samples per intervention set.
Hardware Specification Yes All experiments were executed on a machine equipped with an Apple M2 processor and 8GB of RAM.
Software Dependencies No We implement q NEHVI using the botorch library.
Experiment Setup Yes The batch size is set to 5. For reproducibility, all experiments are run across 10 random seeds, resulting in varying initializations of D. ... Par EGO: ...σ = 0.5 as initial standard deviation. ... TSEMO: ...use 100 points for spectral sampling. ... q NEHVI: ...use 10 optimization restarts, and 64 raw samples for acquisition maximization. Moreover, the acquisition function uses a Sobol QMC sampler with 128 samples.