원문정보
Survey on Multitask Reinforcement Learning for Engineering Design
초록
영어
Engineering design problems rely mainly on human knowledge. So it is difficult to present creative designs and has limitations that they do not deviate from certain design patterns. Machine learning has been suggested as a solution to address these problems. Therefore, there are ongoing efforts to apply machine learning to engineering design. Unlike supervised learning, which learns based on correct answers, reinforcement learning finds good action through trial and error without prior knowledge. So it is possible to find new design methods that are different from existing practices. In this paper we introduce a multitask reinforcement learning that has been modified to handle different design goals from existing reinforcement learning, and then introduce some case studies that applied reinforcement learning to engineering design problems.
목차
1. Introduction
2. Preliminaries
3. Multitask Rinforcement Learning
3.1. Universal Value Function Approximators
3.2. Hindsight Experience Replay
4. Case Study of Engineering Design Application
4.1. Single Task
4.2. Multitask Case: Inertial Flow Sculpting
5. Conclusions
Acknowledgement
References