Efficient optimization of large probabilistic models

• Combination of genetic algorithms with artificial neural networks for speedup.
• Optimization of probabilistic system models.
• Evaluation on a real world case-study from the railroad domain.
• Using quantitative safety analysis as optimization goal.

The development of safety critical systems often requires design decisions which influence not only dependability, but also other properties which are often even antagonistic to dependability, e.g., cost. Finding good compromises considering different goals while at the same time guaranteeing sufficiently high safety of a system is a very difficult task.We propose an integrated approach for modeling, analysis and optimization of safety critical systems. It is fully automated with an implementation based on the Eclipse platform. The approach is tool-independent, different analysis tools can be used and there exists an API for the integration of different optimization and estimation algorithms. For safety critical systems, a very important criterion is the hazard occurrence probability, whose computation can be quite costly. Therefore we also provide means to speed up optimization by devising different combinations of stochastic estimators and illustrate how they can be integrated into the approach.We illustrate the approach on relevant case-studies and provide experimental details to validate its effectiveness and applicability.