Active Learning

Standard Bayesian Optimization evaluates one point per iteration, while batch Bayesian Optimization selects multiple points at once. The framework used active learning and multi-objective Bayesian optimization to choose informative experiments sequentially. The KadiAIgent plugin implemented Bayesian optimization and communicated with a Bayesian inference service. The active-learning method used independent Gaussian process surrogate models for each objective. The acquisition strategy used q-noisy Expected Hypervolume Improvement to handle noisy measurements and batch optimization. Batch Gaussian Process UCB requires diverse intra-batch exploration for bounded hallucinated information and sublinear regret. Naively selecting the top acquisition maximizers in batch Bayesian Optimization tends to produce nearly identical candidates. Batches 11 through 17 were generated by the active-learning model after manually defined initial batches. Bayesian Optimization is used to optimize expensive black-box functions.