Semantic parsing of automobile steering systems

Formal specification plays crucial roles in the rigorous verification and design of automobile steering systems. The challenge of getting high-quality formal specifications is well documented. This paper presents a problem called 'semantic parsing', the goal of which is to automatically translate the behavior of an automobile steering system to a formal specification written in signal temporal logic (STL) with human-in-the loop manner. To tackle the combinatorial explosion inherent to the problem, this paper adopts a search strategy called agenda-based parsing, which is inspired by natural language processing. Based on such a strategy, the semantic parsing problem can be formulated as a Markov decision process (MDP) and then solved using reinforcement learning. The obtained formal specification can be viewed as an interpretable classifier, which, on the one hand, can classify desirable and undesirable behaviors, and, on the other hand, is expressed in a human-understandable form. The performance of the proposed method is demonstrated with study.