首页|期刊导航|Journal of Automation and Intelligence|From black box to physically interpretable:Trustworthy computing for AI-driven decision-making and control

From black box to physically interpretable:Trustworthy computing for AI-driven decision-making and controlOA

中文摘要

As Artificial Intelligence(AI)becomes deeply integrated into the physical world,ensuring verifiable,robust,and safe decision-making and control is paramount.Purely data-driven methods,while successful in perception tasks,often suffer from a‘‘black box’’nature when deployed in safety-critical environments,leading to insufficient generalization and potential hazards.Therefore,the paradigm of trustworthy decision-making and control guided by physical world information has emerged.It integrates physical laws,system dynamics,and environmental constraints into AI models,either as regularization terms in the loss function,embedded within the network architecture,or through hybrid models.It allows AI to simultaneously learn from data and consistently follow physical laws.This paper compares trustworthy computing guided by physical knowledge with traditional data-driven methods,focusing on two representative paradigms:Physics-Informed Neural Networks(PINNs)and Physics-Informed Reinforcement Learning(PIRL).We present their core concepts and integration mechanisms,embedding physical laws via loss regularization or network architecture,and incorporating physics into the learning or control process.This paper reveals the significant improvements of the trustworthy computing in interpretability,robustness,and safety,providing a new path from‘‘black box’’to‘‘physically interpretability’’transformation for safety-critical domains such as autonomous driving,robot collaboration,and industrial process control.Understanding the differences and connections within this paradigm is essential for advancing AI technologies that can serve human society safely and reliably.

Yongduan Song;Anguo Zhang

School of Automation,Chongqing University,Chongqing,400030,ChinaInterdisciplinary Institute for Medical Engineering,Fuzhous University,Fuzhou,350108,China

信息技术与安全科学

Physics-Informed Neural NetworksPhysics-Informed Reinforcement LearningTrustworthy AISafety-critical systems

《Journal of Automation and Intelligence》 2026 (2)

P.91-111,F0002,22

supported in part by the National Key Research and Development Program of China under Grant 2022YFB4701400/4701401in part by the Fundamental Research Funds for the Central Universities,China under Project 2024CDJYXTD007in part by the Graduate Research and Innovation Foundation of Chongqing China under Grant CYS20069in part by the National Natural Science Foundation of China under Grant 62306001,62403082,and 62250710167in part by Natural Science Foundation of Fujian Province,China under grant 2025J01290in part by Chongqing Top-Notch Young Talents Project under Grant cstc2024ycjhbgzxm0085in part by the Natural Science Foundation of Chongqing,China under Grant CSTB2023NSCQ-LZX0026.

10.1016/j.jai.2025.09.003

评论