Core Concepts
A deep reinforcement learning-based framework that autonomously configures and adapts the exploration-exploitation tradeoff throughout the evolutionary computation search process, enhancing the performance of various evolutionary computation algorithms.
Abstract
The paper proposes a deep reinforcement learning-based framework called GLEET (Generalizable Learning-based Exploration-Exploitation Tradeoff) that automatically configures and adapts the exploration-exploitation tradeoff (EET) for evolutionary computation (EC) algorithms.
The key highlights are:
GLEET formulates the EET control as a Markov Decision Process (MDP) and designs a comprehensive state representation, action space, and reward function to enable efficient learning of generalizable EET strategies.
GLEET employs a Transformer-based network architecture that consists of a feature embedding module, a fully informed encoder, and an exploration-exploitation decoder. This allows the network to effectively extract and process the features of the EET and problem knowledge, and adaptively attend to the knowledge of different individuals in the population.
GLEET is evaluated on several representative EC algorithms, including vanilla Particle Swarm Optimization (PSO) and DMSPSO, and demonstrates significant performance improvements over the backbone algorithms as well as existing adaptive and learning-based methods.
GLEET exhibits strong generalization capabilities, performing well on unseen problem classes, dimensions, and population sizes, without the need for retraining.
The learned EET behaviors of GLEET are visualized and interpreted, providing insights into how it learns different strategies for different problems.
Stats
The paper reports the following key metrics:
Optimization performance (mean and standard deviation) of various PSO algorithms on the augmented CEC2021 benchmark problems.
Average rank of the algorithms across the test problems.
Percentage improvement of GLEET-PSO and GLEET-DMSPSO over their backbone algorithms.
Quotes
"GLEET could significantly ameliorate the backbone algorithms, making them surpass adaptive methods and existing learning-based methods."
"GLEET exhibits promising generalization capabilities across different problem classes."