Improving Transferability of Population-Based Learning-to-Optimize Algorithms Through Imitation Learning
Gabriel Anunciação Kopte, Marcelo Gomes Pereira de Lacerda, Fernando Buarque de Lima Neto · 2023
The process of algorithm development is iterative and requires a meticulous evaluation of the existing problem. In recent years, Learning to optimize (L2O) models have transformed this task by using machine learning to generate efficient optimization methods. However, transferring knowledge effectively to problems divergent from the original training context remains a challenge encountered by most L2O approaches. This paper introduces a method that enhances transferability of population-based L2O algorithms through imitation learning. Our method adopts a two-step learning process involving imitation and reinforcement learning to train L2O policies. Experimental results demonstrate that our proposed model, incorporating Particle Swarm optimization (PSO) as the teacher algorithm, outperforms the baseline model in key scenarios. This validates the potential of imitation learning in catalyzing the progress of populations-based L2O models.