Adaptive Multi-Objective Deep Reinforcement Learning for Policy-Driven Neural Architecture Search of Resource-Efficient CNNs

Abstract

Neural Architecture Search (NAS) automates the process of designing neural networks but the focus is mostly on prediction accuracy without considering the model complexity and computational cost. This research study proposes a new framework called Adaptive Multi-Objective Policy-Driven Neural Architecture Search (AMO-PDNAS) based on Deep Reinforcement Learning (DRL). The CNN architecture design is formulated using Markov Decision Process. A parameterized policy network generates CNN architecture parameters, which include kernel size, number of filters, activation functions, and dropout rate. Reward weighting is adapted through each search cycle in an attempt to shift from an explorative approach driven by accuracy towards an approach driven by both accuracy and efficiency considerations. Candidate architectures are learned and evaluated in cycles and feedback is provided to the policy based on the REINFORCE algorithm. Results on the MNIST problem show robust behavior of searching for high-accuracy architectures along with efficient trade-offs of accuracy and complexity.