Max-value Entropy Search for Efficient Bayesian Optimization

About

Entropy Search (ES) and Predictive Entropy Search (PES) are popular and empirically successful Bayesian Optimization techniques. Both rely on a compelling information-theoretic motivation, and maximize the information gained about the $\arg\max$ of the unknown function; yet, both are plagued by the expensive computation for estimating entropies. We propose a new criterion, Max-value Entropy Search (MES), that instead uses the information about the maximum function value. We show relations of MES to other Bayesian optimization methods, and establish a regret bound. We observe that MES maintains or improves the good empirical performance of ES/PES, while tremendously lightening the computational burden. In particular, MES is much more robust to the number of samples used for computing the entropy, and hence more efficient for higher dimensional problems.

Zi Wang, Stefanie Jegelka · 2017

Related benchmarks

Task	Dataset	Result
Bayesian Optimization	50 optimization problems COCO, BoTorch, Bayesmark (aggregated)	Mean RP2.8	26
Simple Regret Minimization	PM2.5 Step 60	25th Percentile Simple Regret11	7
Simple Regret Minimization	Oil Step 30	25th Percentile Simple Regret1.95	7
Simple Regret Minimization	Electrical Grid Stability Step 60	25th Percentile Simple Regret1.49	7
Simple Regret Minimization	PM2.5 Step 120	25th Percentile Simple Regret8	7
Simple Regret Minimization	Asteroid Step 30	25th Percentile Regret0	7
Simple Regret Minimization	Electrical Grid Stability Step 120	Simple Regret (25th Pctl)0	7
Simple Regret Minimization	HPOBench XGB Step 30	25th Percentile Simple Regret0	7
Simple Regret Minimization	HPOBench XGB Step 60	25th Percentile Simple Regret0	7
Simple Regret Minimization	Oil Step 60	25th Percentile Simple Regret0	7