Scholarlink

The autonomous control of aircraft poses a challenging high-dimensional continuous control problem, extensively utilized in drones, autonomous driving systems, and flight simulators. Reinforcement Learning (RL), especially with the latest advancements in deep neural networks, empowers agents to excel in complex continuous control tasks. The research employed the RL framework to model aircraft heading and altitude control as a Markov decision process, simulated via JSBSim. Leveraging the Proximal Policy Optimization (PPO) algorithm alongside a tailored reward function, the agent adeptly learns to attain target altitude and heading at a designated speed through training with an exhaustive set of state variables. Efforts were made to trim down the number of state variables, pinpointing a refined set highly linked to altitude control. This streamlined selection maintains performance while slashing computational complexity and boosting convergence speed. It offers a potent approach for autonomous aircraft control with limited state information.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Autonomous Control of Simulated Aircraft Based on Deep Reinforcement Learning Driven by Limited State Information

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xinglong Shen,
  • Yang Liu,
  • Dequn Zhao

摘要

The autonomous control of aircraft poses a challenging high-dimensional continuous control problem, extensively utilized in drones, autonomous driving systems, and flight simulators. Reinforcement Learning (RL), especially with the latest advancements in deep neural networks, empowers agents to excel in complex continuous control tasks. The research employed the RL framework to model aircraft heading and altitude control as a Markov decision process, simulated via JSBSim. Leveraging the Proximal Policy Optimization (PPO) algorithm alongside a tailored reward function, the agent adeptly learns to attain target altitude and heading at a designated speed through training with an exhaustive set of state variables. Efforts were made to trim down the number of state variables, pinpointing a refined set highly linked to altitude control. This streamlined selection maintains performance while slashing computational complexity and boosting convergence speed. It offers a potent approach for autonomous aircraft control with limited state information.