核心概念
本文介紹一種將形式化驗證技術與基於 AI 的感知系統和經過驗證的控制相結合的方法,以提高自主列車系統的安全性。
統計資料
使用 500 次模擬運行進行蒙地卡羅模擬。
列車在沒有安全防護和經過驗證的控制的情況下,在第一個錯誤檢測到的停車信號前總是提前停車。
經過驗證的控制幾乎消除了所有誤報。
經過驗證的控制導致正確檢測顯著減少。
引述
"This work deals with systems that employ an AI perception system, such as image recognition. Those systems include autonomous vehicles and autonomous railway systems."
"This work presents a real-time demonstrator linking the formal model, the AI, and certified control, extending an approach used to validate reinforcement learning agents [34]."
"The main benefit of our methodology is that we can link the execution of a real AI model to a formal model and check its behaviour using formal properties and statistical validation techniques."