Resource-Aware Neural Network Pruning Using Constrained Reinforcement Learning and Self-Competition

This paper presents a novel approach to neural network pruning by integrating a self-competition reward system into an Automated Machine Learning (AutoML) pruning framework. Traditional pruning methods often rely on handcrafted rules, which can be inefficient and suboptimal. To address these limitations, we propose employing a Constrained Markov Decision Process (CMDP) to model the pruning environment. Our self-competition reward system incentivizes the agent to continuously improve upon its past performance while prioritizing the satisfaction of resource constraints, such as achieving a target compression rate. We demonstrate the effectiveness of our approach through experiments on various datasets, including CIFAR-10/100 and ImageNet, comparing it against state-of-theart pruning methods. The results indicate that our method not only meets the defined compression constraints but also achieves competitive accuracy performance, showcasing an effective and efficient approach to constrained neural network pruning.