7 Maximizing throughput with batch optimization

The BayesOpt loop we have been working with thus far takes in one query at a time and returns the function evaluation for the query before the next query is made. We use this loop in settings where function evaluations can only be made sequentially. However, many real-world scenarios of black box optimization allow the user to evaluate the objective functions in batches. For example, when tuning the hyperparameters of an ML model, we can try out different hyperparameter combinations in parallel if we have access to multiple processing units or computers, instead of running individual combinations one by one. By taking advantage of all the resources available to us, we can increase the number of experiments we conduct and maximize throughput during the function evaluation step of the BayesOpt loop.

7.1 Making multiple function evaluations simultaneously

7.1.1 Making use of all available resources in parallel

7.1.2 Why can’t we use regular BayesOpt policies in the batch setting?

7.2 Computing the improvement and upper confidence bound of a batch of points

7.2.1 Extending optimization heuristics to the batch setting

7.2.2 Implementing batch improvement and UCB policies

7.3 Exercise 1: Extending TS to the batch setting via resampling

7.4 Computing the value of a batch of points using information theory

7.4.1 Finding the most informative batch of points with cyclic refinement

7.4.2 Implementing batch entropy search with BoTorch

7.5 Exercise 2: Optimizing airplane designs

Summary