Minimum Regret Search for Single- and Multi-Task Optimization

2016/02/02 by Jan Hendrik Metzen, Metzen, Jan Hendrik · 1 citation
Computer Science · Decision Sciences · Mathematics · #Advanced Bandit Algorithms Research #Gaussian Processes and Bayesian Inference #Reinforcement Learning in Robotics #cs.IT #cs.LG #cs.RO #math.IT #stat.ML

paper · pdf · doi:10.48550/arxiv.1602.01064

Final version for ICML 2016

arxiv created 2016/05/24 · arxiv updated 2016/05/25

Abstract

We propose minimum regret search (MRS), a novel acquisition function for Bayesian optimization. MRS bears similarities with information-theoretic approaches such as entropy search (ES). However, while ES aims in each query at maximizing the information gain with respect to the global maximum, MRS aims at minimizing the expected simple regret of its ultimate recommendation for the optimum. While empirically ES and MRS perform similar in most of the cases, MRS produces fewer outliers with high simple regret than ES. We provide empirical results both for a synthetic single-task optimization problem as well as for a simulated multi-task robotic control problem.

Citations

• Lipschitzian optimization without the Lipschitz constant
• Taking the Human Out of the Loop: A Review of Bayesian Optimization

Cited by

• Computationally Efficient High-Dimensional Bayesian Optimization via Variable Selection

Related

• Regret-Aware Black-Box Optimization with Natural Gradients, Trust-Regions and Entropy Control
• Direct Regret Optimization in Bayesian Optimization
• Regret Minimization in Scalar, Static, Non-linear Optimization Problems
• MUMBO: MUlti-task Max-value Bayesian Optimization
• Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds