A DQN-Based Reinforcement Learning Model for Neural Network Architecture Search

• Main Authors: Tzu-Han Tsai, 蔡子涵
• Other Authors: Yi-Cheng Chen
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/qf2jk4

碩士 === 國立中央大學 === 資訊管理學系 === 107 === Designing neural network (NN) architectures requires both human expertise and labor. New architectures are handcrafted by careful experimentation or modified from a handful of existing networks. We introduce HCM, a meta-modeling algorithm based on reinforcement learning to automatically generate high-performing NN architectures for a given learning task. The learning agent is trained to sequentially choose NN layers using DQN with an ɛ-greedy exploration strategy and experience replay. The agent explores a large but finite space of possible architectures and iteratively discovers designs with improved performance on the learning task. Even on image classification benchmarks, the agent-designed networks can do good as existing networks designed but more efficient. We also outperform existing meta-modeling approaches for network design on image classification or regression tasks.