关键词: 互联网模式, 云原生, 强化学习, B/S架构, 组件化建模

Abstract:

To solve the problems faced by the adversarial competition mode of agents, including difficult development and deployment, low resource utilization, poor reusability, and difficulty in accessing reinforcement learning algorithms, a new agent simulation training platform was designed. The software components of the competition platform were decoupled based on cloud-native technology; a high-performance simulation engine for the competition environment was proposed; a new method of an embedded reinforcement learning model for an intelligent control terminal was designed, with multiple online and offline policy-based reinforcement learning algorithms set.The experiment demonstrates that the development and deployment of the system is efficient, which reduces the requirements for the hardware equipment of the participants, and the participants can log in to the system via one click without installing any applications. The system runs efficiently and effectively solves problems such as weak concurrency, poor reliability, and response delay through cloud service design. The system is highly adapted to the intelligent training algorithm and supports users to adjust the training, deduction modules, and main parameters, so as to reduce the threshold of intelligent training for participants.

Key words: internet mode, cloud-native, RL, B/S architecture, component-based modeling