On the Effectiveness of Signal Rescaling in Hybrid System Falsification

Hybrid system falsification employs stochastic optimization to search for counterexamples to a system specification in Signal Temporal Logic (STL), guided by quantitative STL robustness. The scale problem could arise when the STL formula is composed of sub-formulas concerning signals having different scales (e.g., speed [km/h] and rpm): the performance of falsification could be negatively affected because different scales can mask each other’s contribution to robustness. A natural solution consists in rescaling the signals to the same order of magnitude. In this paper, we investigate whether this “basic” approach is always effective, or better rescaling strategies could be devised. Experimental results show that basic rescaling is not always the best strategy, and sometimes “unbalanced” rescalings work better. We investigate the reasons of this, and we identify future research directions based on this observation.