| Title |
NAS based on Reinforcement Learning with Improved Multi-objective Reward Function |
| Authors |
임철순(Cheolsun Lim) ; 김명선(Myungsun Kim) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.11.39 |
| Keywords |
Neural architecture search; DNN architecture; Mobile; Reinforcement learning; Embedded system |
| Abstract |
As the utility of DNNs is verified in a wide range of applications, many research works have been conducted to lighten DNN models for the use in embedded environments such as mobiles. However, since it takes lots of effort and time to design a light-weight DNN structure directly, NAS (Neural Architecture Search) has been introduced to figure out this problem. NAS automatically explores the DNN architectures, where the explored DNNs have the execution performances that are not inferior to the existing state-of-the-art (SOTA) DNNs. MnasNet, a typical mobile NAS, can explore DNNs that satisfy high performance and low execution time. The goal of MnasNet is to explore DNNs with the highest accuracy while satisfying the inference execution time constraints. In this paper, we focus on a multi-objective reward function that optimizes for the accuracy and the inference time based on MnasNet. First, we analyze the reward function used by MnasNet and propose a new reward function to explore DNNs with the maximum accuracy while having a reduced inference execution time than a given time limit. When we use the reward function proposed in this paper, the point at which the DNN models generated by NAS satisfy the inference execution time constraint is 15% faster than when using the reward function of MnasNet. |