A Reinforcement Learning Approach to Design Verification Strategies of Engineered Systems
Peng Xu, Alejandro Salado, Guangrui Xie · 2020
System verification is a critical process in the development of engineered systems. Engineers gain confidence in the correct functionality of the system before it is deployed into operation by executing verification activities. Choosing the right set of verification activities at the right system development stage, that is, designing a verification strategy (VS), is essential to balancing information discovery and verification cost. Only recently, quantitative methods have been proposed to support the design of verification strategies. However, their applicability in real-life scenarios is impractical due to their limited computational efficiency in the high dimensional solution space of the VS selection problem. This paper presents a reinforcement learning (RL) approach to search for a near-optimal VS. Specifically, the VS design problem is formulated as a Markov decision process (MDP) in which a value function is required. Then we combine tree search and a neural network (NN) to design a RL algorithm. In the RL algorithm, the value function is approximated as a NN that is trained in an iterative way. The near-optimal VS can be generated from the trained NN. A case study is presented to show the superiority of the proposed method.