AutoMH: Automatically Create Evolutionary Metaheuristic Algorithms Using Reinforcement Learning

This research aims to have a system that can automatically create-discover evolutionary metaheuristic algorithms. This system was called the AutoMH framework. The AutoMH framework is based on Reinforcement Learning in which various improvements have been made to its architecture. Mainly the concept of reward, which in this research it works by ranking the generated algorithms.

Initially, AutoMH to create metaheuristic algorithms starts with a base algorithm structure. Later AutoMH makes various modifications to the design of these metaheuristic algorithms. These algorithms must solve a portfolio of problems and are subsequently ranked by efficiency.

The experiments were performed on an P65 Creator 9SE laptop with Intel Core I7 9750H CPUs and 16 GB RAM. In approximately 1 hour, the AutoMH found a metaheuristic algorithm that solves the portfolio of problems. We will call this found algorithm AMH.

The AMH algorithm designed by AutoMH was compared to 14 metaheuristic algorithms: BAT, CS, DE, FFA, GA, GWO, HHO, JAYA, MFO, MVO, PSO, SCA, SSA, and WOA; where the AMH algorithm has obtained ranking 1 and 2 by solving 8 and 4 optimisation problems, respectively.

If we look at Search Trajectory Networks, the AMH metaheuristic algorithm designed by AutoMH adapts to the portfolio of problems, finding a solution faster and more accurately.

The development of this research was financed with personal funds. While the APC of the scientific article was financed with 18 discount vouchers granted for performing peer review, and the remainder of the APC was paid with personal funds.

Citation:

Almonacid, B. (2022). AutoMH: Automatically Create Evolutionary Metaheuristic Algorithms Using Reinforcement Learning. Entropy, 24(7), 957. https://doi.org/10.3390/e2407095