Abstract:
|
One of the most challenging tasks in automated virtual testing of complex systems is the failure analysis. An essential part here is to learn the boundary between pass and fail test results in a high dimensional input space, which we call system boundary. In order to estimate the system boundary representing a manifold in a high-dimensional input space, we compare the usage of neural networks and gaussian processes as predictors. Both algorithms are known to be very flexible, however gaussian processes inherently include an uncertainty estimate for the prediction. These predictors are then fed into a sampling schema, which is refining the estimate of the system boundary. In a follow up task, clustering methods (e.g. hierarchical clustering) are used to estimate various error regimes. There are two main advantages of the presented approach: 1) simulation workload can be reduced by focussing virtual testing close to the system boundary and 2) detailed information about system performance changes and different areas of weak system performance can be provided to software development teams. The approach was applied to a virtual testing use case for autonomous vehicles.
|