An Actor-Critic Method for Simulation-Based Optimization

In this work, we study simulation-based optimization, where the agent aims to select best configuration from design space with as few possible iterations. Inspired by success of deep reinforcement learning (DRL), formulate sampling process policy searching and give a solving method perspective iteration. Concretely, surrogate model for predicting performance each parameterized are applied, which correspond critic actor in actor-critic (AC) method, respectively. We further derive updating rule propose two algorithms selection continuous discrete spaces, Finally, validated experimentally on 1) toy examples intuitively explain principle 2) high-dimensional tasks reveal effectiveness large-scale problems. The results show that proposed can efficiently deal problems effectively eliminate sub-optimal configurations.