Asynchronous Parallel Surrogate Optimization Algorithm for Quantitative Strategy Parameter Tuning

摘要

Surrogate-model based optimization algorithms can be applied to solve expensive black-box function optimization problem. With the introduction of ensemble model, surrogate-model based algorithms can be automatically adjusted to adapt to various specific problems with different parameter spaces and no need for manual design of surrogate model. However, introduction of ensemble model significantly increases the computational load of surrogate-model based algorithms for training and updating of ensemble model. In this article, parallel computing technology is utilized to speed up the weight updating related computation for the ensemble surrogate model built by Dempster-Shafer theory, and a novel parallel sampling mechanism based on stochastic response surface method is developed to implement asynchronous parameter optimization, based on witch an asynchronous parallel global optimization algorithm is proposed. Furthermore, the parallel algorithm proposed is applied to quantitative trading strategy tuning in financial market and shows both feasibility and effectiveness in actual application. Experiments demonstrates that, the algorithms can achieve high speedup ratio and scalability with no degradation of optimization performance.