Evolutionary Sampling Agent for Expensive Problems

Data-driven evolutionary algorithms are widely studied for their ability to solve expensive optimization problems in engineering and science. This article introduces a novel optimization framework to solve costly optimization problems, called the evolutionary sampling agent (ESA). ESA considers the optimization algorithm as an agent, which operates on four different characteristics of evolutionary sampling strategies to search the global optimum. Among these four evolutionary sampling strategies, the first strategy prefers exploration, the second and the fourth strategies use different local search methods preferring exploitation, and the third strategy integrates good genes from historical solutions. ESA consists of two layers of learning mechanisms. On the one hand, the evolutionary sampling strategies use historical data to construct surrogate models to efficiently sample a candidate solution. On the other hand, the agent adjusts the probability of selecting different sampling strategies through the feedback information received in the optimization process. Seven benchmark functions with 30, 50, and 100 dimensions were adopted. Compared with the other state-of-the-art methods, the results show that ESA yields a promising performance for expensive problems.

Automated summary

This article introduces a novel optimization framework called the evolutionary sampling agent (ESA), which treats the optimization algorithm as an agent and utilizes four different evolutionary sampling strategies to search the global optimum. ESA consists of two layers of learning mechanisms, using surrogate models constructed from historical data to efficiently sample candidate solutions and adjusting the probability of selecting different sampling strategies through feedback information. Experimental results demonstrate that ESA performs well for solving expensive problems compared to other state-of-the-art methods.