A Multi-Scaled Method for Parallel Bayesian Optimization in Deep Predictive Analytics

This paper proposes a new roll-out scaling method for parallel Bayesian optimization and discusses how the proposed multi-scaled optimization guarantees a better convergence speed with outperformed accuracy than the conventional parallel search algorithms. Experiment results demonstrate that an entire search space can be efficiently reduced to more feasible subdomains. The performance of parallel Bayesian search can be further accelerated based on the interchangeable local evidence by properly adjusting three quantitative aspects in terms of space factorization, search direction, and architecture scaling.