Machine learning model selection with multi-objective Bayesian optimization and reinforcement learning

A machine learning system, including when used in reinforcement learning, is usually fed with only limited data, while aimed at training a model with good predictive performance that can generalize to an underlying data distribution. Within certain hypothesis classes, model selection chooses a model based on selection criteria calculated from available data, which usually serve as estimators of generalization performance of the model. One major challenge for model selection that has drawn increasing attention is the discrepancy between the data distribution where training data is sampled from and the data distribution at deployment. The model can over-fit in the training distribution, and fail to extrapolate in unseen deployment distributions, which can greatly harm the reliability of a machine learning system. Such a distribution shift challenge can become even more pronounced in high-dimensional data types like gene expression data, functional data and image data, especially in a decentralized learning scenario. Another challenge for model selection is efficient search in the hypothesis space. Since training a machine learning model usually takes a fair amount of resources, searching for an appropriate model with favorable configurations is by inheritance an expensive process, thus calling for efficient optimization algorithms. To tackle the challenge of distribution shift, novel resampling methods for the evaluation of robustness of neural network was proposed, as well as a domain generalization method using multi-objective bayesian optimization in decentralized learning scenario and variational inference in a domain unsupervised manner. To tackle the expensive model search problem, combining bayesian optimization and reinforcement learning in an interleaved manner was proposed for efficient search in a hierarchical conditional configuration space. Additionally, the effectiveness of using multi-objective bayesian optimization for model search in a decentralized learning scenarios was proposed and verified. A model selection perspective to reinforcement learning was proposed with associated contributions in tackling the problem of exploration in high dimensional state action spaces and sparse reward. Connections between statistical inference and control was summarized. Additionally, contributions in open source software development in related machine learning sub-topics like feature selection and functional data analysis with advanced tuning method and abundant benchmarking were also made