Testing Adaptive Software with Probabilistic Guarantees
Martina Maggio, Saarland University, Germany, and Lund University, Sweden
Testing software that adapts, like a machine learning algorithm, is very complicated. In most cases, it is very difficult - if not impossible - to conduct exhaustive testing and analyse each possibile configuration. This is not only because the space of the configurations is very large, but also because the software learns and adapts, and running the same function with the same set of inputs may result in different outcomes. Martina Maggio will speak of this matter in a virtual seminar November 25 at 17:30 GMT.
Virtual link: https://us02web.zoom.us/j/85745915221
In this context, it is impossible to get a deterministic answer to the software correctness, and there is a need for a paradigm shift to the probabilistic setup. In our research, we explored different alternatives to obtain probabilistic guarantees. The classical tools from statistics are Monte Carlo simulations and the Extreme Value Theory. We show that these tools have limitations that can be overcame by formulating the problem of testing a software that adapts itself as a chance-contrained optimization problem. In doing so, we employ the scenario theory, from the field of robust control.
This seminar is part of the joint ICE-TCS/GSSI Virtual seminars.
