Continuous Parameter Control in Genetic Algorithms using Policy Gradient Reinforcement Learning

Alejandro de Miguel Gomez, Farshad Ghassemi Toosi

2021

Abstract

Genetic Algorithms are biological-inspired optimization techniques that are able to solve complex problems by evolving candidate solutions in the search space. Their evolutionary features rely on parameterized stochastic operators that are sensitive to changes and, ultimately, determine the performance of the algorithms. In recent years, Reinforcement Learning has been proposed for online parameter control in contrast to traditional fine-tuning, which inevitably leads to suboptimal configurations found through extensive trial-and-error. In this regard, the current literature has focused on value-based Reinforcement Learning controllers for Genetic Algorithms without exploring the advantages of policy gradient methods in such environments. In this study, we propose a novel approach to leverage the continuous nature of the latter with agents that learn a behavior policy and enhance the performance of Genetic Algorithms by tuning their operators dynamically at runtime. In particular, we look at Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO). The resulting hybrid algorithms are tested on benchmark combinatorial problems and performance metrics are discussed in great detail considering the existing work based on Q-Learning and SARSA.

Paper Citation

in Harvard Style

de Miguel Gomez A. and Toosi F. (2021). Continuous Parameter Control in Genetic Algorithms using Policy Gradient Reinforcement Learning. In Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: ECTA; ISBN 978-989-758-534-0, SciTePress, pages 115-122. DOI: 10.5220/0010643500003063

in Bibtex Style

@conference{ijcci21,
author={Alejandro de Miguel Gomez and Farshad Ghassemi Toosi},
title={Continuous Parameter Control in Genetic Algorithms using Policy Gradient Reinforcement Learning},
booktitle={Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: ECTA},
year={2021},
pages={115-122},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010643500003063},
isbn={978-989-758-534-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: ECTA
TI - Continuous Parameter Control in Genetic Algorithms using Policy Gradient Reinforcement Learning
SN - 978-989-758-534-0
AU - de Miguel Gomez A.
AU - Toosi F.
PY - 2021
SP - 115
EP - 122
DO - 10.5220/0010643500003063
PB - SciTePress