Deep Mean-Kernel Learning for Bayesian Optimization in IoT-Enabled Data Center

In efforts to mitigate energy consumption in large data centers, numerous optimization strategies have been explored. However, these methods often encounter challenges including prolonged optimization times and settling on local optima. Among diverse algorithms, Bayesian optimization (BO) demonstrates exceptional performance. Due to the slow convergence and limited scalability of existing Bayesian methods in handling high-dimensional data, we explore the influence of the surrogate model on BO, and introduce an innovative optimization approach deep mean-kernel learning, which can converge rapidly to the global optimal solution in high dimensional space and is specifically designed for the refrigeration systems of data centers. In our approach, a neural network is employed to approximate the conventional mean function, while an infinite-width neural network serves as the kernel function, modifying the Gaussian process. Utilizing two prevalent acquisition functions, we conduct comparative experiments of our algorithm against existing optimization methods in the context of data center scenarios. Our findings demonstrate that our approach adeptly manages high-dimensional data and exhibits rapid convergence towards the global optima.