Bayesian Optimisation for Efficient ML Pipeline Hyperparameter Tuning Under a Cost Budget
Bayesian optimization (BO) offers a simple mechanism of optimising functions whose closed form expression may not be known, within a limited iteration budget. This is particularly useful in the problem of hyperparameter tuning. At each iteration, different regions in a search space of inputs may incur different costs, i.e time units required to perform a query. The function which assigns a cost to each point in the input space may also not be known. Much of the BO literature, however, tends to focus on the number of iterations incurred for convergence, rather than the total costs incurred. Furthermore, little research has been done on specified formulations of BO which are needed for ML pipeline hyperparameter tuning. In the ML pipeline setting, we wish to tune hyperparameters of not just a machine learning model, but also of any additional steps involved in the overall artificial intelligence system. For example, in computer vision, we may be concerned with image preprocessing and output postprocessing steps, and the hyperparameters involved in these stages our pipeline. When performing hyperparameter search in an ML pipeline setting, we can cache hyperparameter combinations from earlier stages of a pipeline, once queried, to be reused with hyperparameter combinations of latter stages in a pipeline, for a more effecient search. This cache-and-reuse is referred to as memoisation. We propose a new acquisition function for BO, referred to as EEIPU, that is uniquely suited to the pipeline setting. Furthermore, we explore techniques regarding the combination of EEIPU, with memoised hyperparameter search. One key feature of EEIPU is that we design it to work in a scenario whereby the closed form of the cost function at each pipeline stage is unknown. In addition, we devote extensive resources to designing ML pipeline tuning systems, which can work hand-in-hand with our acquisition function. These systems are built towards allowing our acquisition function to be deployed in real-world industry or research use cases, for entire-pipeline hyperparameter tuning. We study the experimental behaviour of EEIPU, and compare it with related acquisition functions in the literature, to better understand its performance, and propose future research directions.
M.S. Nwadike, "Bayesian Optimisation for Efficient ML Pipeline Hyperparameter Tuning Under a Cost Budget", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2022.