Generative Evolutionary Strategy For Black-Box Optimization

Changhwi Park

Keywords: Non-convex optimization, black-box optimization

TL;DR: The combination of generative surrogate networks and evolutionary strategies significantly increases the performance of high-dimensional black-box optimization.

Abstract: Numerous scientific and technological challenges arise in the context of optimization, particularly, black-box optimization within high-dimensional spaces presents significant challenges. Recent investigations into neural network-based black-box optimization have shown promising results. However, the effectiveness of these methods in navigating high-dimensional search spaces remains limited. In this study, we propose a black-box optimization method that combines an evolutionary strategy (ES) with a generative surrogate neural network (GSN) model. This integrated model is designed to function in a complementary manner, where ES addresses the instability inherent in surrogate neural network learning associated with GSN models, and GSN improves the mutation efficiency of ES. Based on our experimental findings, this approach outperforms both classical optimization techniques and standalone GSN model

Submission Number: 2790