Due to COVID-19 outbreak across the world, we need to move our theory seminars online to ZOOM. We are staying on track with the schedule and Yoriyuki Yamagata will give his talk tomorrow at 2pm.
Topic: Falsification of Cyber-Physical Systems Using Deep Reinforcement Learning.
Abstract: “Falsification” is a method to find a system input or parameter (counter-example) which causes a behavior violating a given specification (usually given by metric or signal temporal logic). Because the correctness of a complex CPS is difficult to be proven, falsification is more practical approach than full verification. A counter-example found by falsification can be used for debugging and testing. Failure of falsification does not generally mean the correctness of the system, but suggests it in some degree. “Robustness guided falsification” is an approach of falsification. “Robustness” is a numerical measure of how robustly a formula holds. If robustness becomes negative, the formula is false. Therefore, minimizing robustness can lead falsification of a formula.
In this talk, we introduce a method to recast robustness guided falsification to a “reinforcement learning problem”. Reinforcement learning is a machine learning technique in which an agent finds a law of an interacting environment and maximizes a reward. We implement our method using “deep reinforcement leaning”, in which deep neural networks are used, and present a case study to explore its effectiveness. (This work is a collaboration with Shuang Liu, Takumi Akazaki, Yihai Duan, Jianye Hao)