컨텐츠 시작
학술대회/행사
초록검색
| 제출번호(No.) | 0186 |
|---|---|
| 분류(Section) | Special Session |
| 분과(Session) | (SS-20) Mathematical Finance (SS-20) |
| 발표시간(Time) | 20th-C-10:20 -- 10:40 |
| 영문제목 (Title(Eng.)) |
Solving portfolio optimization problem using PINN with policy refinement |
| 저자(Author(s)) |
Jeonggyu Huh1 Sungkyunkwan University1 |
| 초록본문(Abstract) | To optimize a financial portfolio in a continuous-time model, it is necessary to find the optimal investment policy function. However, it is difficult to obtain the policy function directly, so the traditional approach is to solve the Hamilton-Jacobi-Bellman (HJB) partial differential equation to derive the corresponding value function and find the policy function implicitly in the value function. However, this approach is difficult to use when the constraints involved become complex, so it has been proposed to approximate the policy function with a neural network and train the policy network. This approach is good in terms of generality, but has the disadvantage of slow training speed. Therefore, in this study, we aim to improve the training speed of the policy neural network by solving the HJB equation with a physical information neural network (PINN), evaluating the value function, improving the policy neural network according to the re-evaluated value function, and repeating this process until convergence. In this way, we show that this methodology can achieve both training speed and feasibility under a general condition. |
| 분류기호 (MSC number(s)) |
91G10 |
| 키워드(Keyword(s)) | Portfolio optimization, value function, PINN, policy refinement, reinforcement learning |
| 강연 형태 (Language of Session (Talk)) |
Korean |