This package includes an API that connects with a backend that executes our algorithm for black-box optimization. The proposed algorithm is a novel method that enables automatic hyperparameter optimization in very high dimensional spaces. The input to the algorithm is (1) an objective function which is going to be optimized, (2) the underlying free hyperparameters, and (3) a set of constraints for optimality. The algorithm outputs the optimal hyperparameters that maximize or minimize the given objective function while adhering to the provided constraints.
The system consists of an adaptive sampler engine, an evaluation environment, and a scoring oracle. The adaptive sampler engine iteratively generates a set of hyperparameter candidates, dubbed samples. The samples are then evaluated in the evaluation environment to measure their characteristics. Using the obtained evaluations, we use the scoring oracle to assign scores to each sample. The given scores represent the optimality of the samples according to the objective function and the constraints. Based on the scores, the adaptive sampling engine updates its sampling distribution so that the newly generated samples are better than the previous ones. By sequentially repeating the above steps, the adaptive sampler engine is able to locate samples, i.e., hyperparameters, that optimize the desired objective function while adhering to the provided constraints.
The above system has a wide use case for automating design in various fields, e.g., circuit design, compilers, machine learning optimization, to name a few. The software significantly reduces the engineering costs and design time by automating tasks that were formerly performed by human experts through manual trial and error.