Improved Learning for Stochastic Timed Models by State-Merging Algorithms

The construction of faithful system models for quantitative analysis, e.g., performance evaluation, is challenging due to the inherent systems’ complexity and unknown operating conditions. To overcome such difficulties, we are interested in the automated construction of system models by learning from actual execution traces. We focus on the timing aspects of systems that are assumed to be of stochastic nature. In this context, we study a state-merging procedure for learning stochastic timed models and we propose several enhancements at the level of the learned model structure and the underlying algorithms. The results obtained on different examples show a significant improvement of timing accuracy of the learned models.