體驗區
|
| 篇名 |
應用於隨機干擾飛行環境之四軸無人機強化學習控制器設計
|
| 並列篇名 |
Design of Reinforcement Learning Controller for Quadcopter in Flight Environment with Random Disturbance
|
| 作者 |
劉俊宏(Chun-Hung Liu)
、葉舜斌(Shun-Pin Yeh)
、王昱健(Yu-Chien Wang)
、賴威霖(Wei-Lin Lai)
、沈尚錡(Shang-Chi Shen)
、丁澤安(Ze-An Ding)
、朱力民(Li-Ming Chu)
|
| 中文摘要 |
現今多數四軸無人機採用比例-積分-微分(Proportional-Integral-Derivative, PID)控制器，而PID參數常需依靠較為耗時之經驗法則進行調整，且針對特定不同飛行環境(如高度差變化量過大)，原調整後之PID參數亦須再次重新設計與調整，不然將導致PID控制器無法發揮最佳之控制性能。近來研究指出，機器學習領域中之強化學習技術可以解決高複雜度系統問題，透過接收相同飛行環境之不同誤差回饋持續進行學習，藉以獲得最佳化決策。故本研究著重於開發強化學習技術並應用至四軸無人機之姿態控制，並透過使用近端策略優化(Proximal Policy Optimization, PPO)演算法設計無人機之強化學習控制器，提高該控制器之飛行環境適應能力。本強化學習控制器與PID控制器於不同目標高度，將無具有及具有外部隨機干擾之情況進行控制性能分析比較，結果發現強化學習控制器之控制性能(含暫態及穩態響應)皆優於PID控制器，故本研究初步結論為無論在面對有無外部隨機干擾之飛行環境中，訓練完成之強化學習控制器具備更佳之環境適應性與控制能力。
|
| 英文摘要 |
Most quadrotors adopt Proportional-Integral-Derivative (PID) controllers, that requires time-consuming empirical tuning of PID parameters. Moreover, the parameters need to be redesigned and readjusted for various flight environments to achieve optimal control performance, especially when desired altitude is changed. Recently, some studies have demonstrated that reinforcement learning (RL) in the field of machine learning can address highly complex problems for the system. RL adopts continuous learning through feedback of different errors under the same flight environment to obtain the optimal decision for the system. Therefore, this study focuses on developing RL technology and applying it to an attitude control of quadcopters. In addition, a Proximal Policy Optimization (PPO) algorithm is used to design a RL controller for the quadcopter to enhance control performance for various flight environments. The performance of the RL controller is compared to that of the PID controller in various target altitudes without and with adding external random disturbances. The simulation results indicate that the RL controller performs better than the PID controller in terms of control performance, including transient and steady-state responses. This study preliminarily concluded that as compared with the well-tuned PID controller, this well-trained RL controller is able to have better environmental adaptability and control capability in various flight environments.
|
| 頁次 |
045-054
|
| 關鍵詞 |
四軸無人機
、比例-積分-微分控制器
、強化學習控制器
、近端策略優化演算法
、隨機干擾環境
、Quadcopter
、PID controller
、Reinforcement Learning Controller
、Proximal Policy Optimization Algorithm
、Flight Environment with Random Disturbance
|
| 卷期 |
13:1
|
| 日期 |
202305
|
| 刊名 |
臺東大學綠色科學學刊
|
| 出版單位 |
國立臺東大學理工學院
|
|
相關文獻 :
客服信箱