中圖分類號： TP181 文獻(xiàn)標(biāo)識碼： A DOI： 10.19358/j.issn.2096-5133.2022.05.012 引用格式： 劉志飛，曹雷，賴俊，等. 基于多智能體深度強(qiáng)化學(xué)習(xí)的無人機(jī)集群自主決策[J].信息技術(shù)與網(wǎng)絡(luò)安全，2022，41(5)：77-81.
Utonomous decision making of UAV cluster with multi-agent deep reinforcement learning
Li Zhifei，Cao Lei，Lai Jun,Chen Xiliang
(College of Command and Control Engineering，Army Engineering University，Nanjing 210007，China)
Abstract： Because the traditional UAV is controlled manually, UAV cluster is more rigid in the strong electromagnetic interference and complex and changeable battlefield environment. In the study, a flexible and intelligent UAV controller is developed. With a neural network trained by multi-agent deep reinforcement learning technology,UAV can control his behavior in flight. At the same time,UAV obtains state information from the battlefield environment, makes independent decisions, forms an effective combat formation with other UAVs, flexibly coordinates and cooperates with each other, and produces the optimal action.
Key words : unmanned aerial vehicle；reinforcement learning；multi agent；autonomous decisio
0 引言
對人工操縱無人機(jī)來說，同時操控多架無人機(jī)完成多項任務(wù)且無人機(jī)之間形成有效配合是相當(dāng)困難的，注意力分散或者操控失誤都會造成較大的安全風(fēng)險。無人機(jī)的操控還受到電磁干擾和遠(yuǎn)程控制距離的限制，因此，無人機(jī)靈活自主決策能力顯得尤為重要。近年來，多智能體深度強(qiáng)化學(xué)習(xí)(Multi-Agent Deep Reinforcement Learning，MADRL)在復(fù)雜游戲中取得完勝人類專家水平的勝利，表明多智能體深度強(qiáng)化學(xué)習(xí)在解決復(fù)雜序貫問題上取得重要突破。強(qiáng)化學(xué)習(xí)技術(shù)應(yīng)用到無人機(jī)群可以提高無人機(jī)群的靈活智能性。本文以一個由6架無人機(jī)組成的無人機(jī)群為例，使用墨子AI仿真實驗平臺，無人機(jī)群組成一個巨大的動作空間，時間步內(nèi)有200多個組合的動作空間，為每架無人機(jī)在每一步行為的機(jī)動方向、航線或向目標(biāo)發(fā)出攻擊都有提供了上千種選擇。使用深度神經(jīng)網(wǎng)絡(luò)來預(yù)測每個無人機(jī)在每個時間步的最優(yōu)動作，并根據(jù)每個無人機(jī)的局部觀察產(chǎn)生自主決策。MADRL方法生成無人機(jī)群作戰(zhàn)決策對無人機(jī)作戰(zhàn)研究具有重要的參考價值，是未來人工智能應(yīng)用在軍事領(lǐng)域的重要方向。
