Enhancing Temporal Logic Falsification of Cyber-Physical Systems using multiple objective functions and a new optimization method

Licentiate thesis, 2020

Cyber-physical systems (CPSs) are engineering systems that bridge the cyber-world of communications and computing with the physical world. These systems are usually safety-critical and exhibit both discrete and continuous dynamics that may have complex behavior. Typically, these systems have to satisfy given specifications, i.e., properties that define the valid behavior. One commonly used approach to evaluate the correctness of CPSs is testing. The main aim of testing is to detect if there are situations that may falsify the specifications.

 

For many industrial applications, it is only possible to simulate the system under test because mathematical models do not exist, thus formal verification is not a viable option. Falsification is a strategy that can be used for testing CPSs as long as the system can be simulated and formal specifications exist. Falsification attempts to find counterexamples, in the form of input signals and parameters, that violate the specifications of the system. Random search or optimization can be used for the falsification process. In the case of an optimization-based approach, a quantitative semantics is needed to associate a simulation with a measure of the distance to a specification being falsified. This measure is used to guide the search in a direction that is more likely to falsify a specification, if possible.

 

The measure can be defined in different ways. In this thesis, we evaluate different quantitative semantics that can be used to define this measure. The efficiency of the falsification can be affected by both the quantitative semantics used and the choice of the optimization method. The presented work attempts to improve the efficiency of the falsification process by suggesting to use multiple quantitative semantics, as well as a new optimization method. The use of different quantitative semantics and the new optimization method have been evaluated on standard benchmark problems. We show that the proposed methods improve the efficiency of the falsification process.