Our paper Search-Based Testing of Reinforcement Learning has been accepted to IJCAI 2022.
In the paper, written together with Filip Cano, Bernhard K. Aichernig, and Bettina Könighofer, we propose a search-based testing framework for testing (deep) reinforcement learning agents with respect to safety and robustness.
We received a best paper award at the 19th International Conference on Software Engineering and Formal Methods for our paper entitled “Active Model Learning of Stochastic Reactive Systems“. In the paper, we present an L*-based algorithm for learning stochastic Mealy machines, which we show to be more efficient than learning of Markov decision processes.
The implementation of our algorithm is available in AALpy, our active automata learning library written in Python. So, if you want to experiment with the algorithm yourself, check out AALpy on GitHub.
I have joined the Dependable Embedded Systems Lab (DES Lab) as a postdoc. The DES Lab is a collaboration of Graz University of Technology and Silicon Austria Labs (SAL). I am excited and looking forward to working in an interdisciplinary team of researchers with expertise spanning the areas of verification, testing, machine learning, security, and embedded systems. Our goal is to make today’s smart embedded systems dependable.
Our paper Online Shielding for Stochastic Systems has been accepted for presentation at NFM 2021.
We perform on-the-fly safety computations to block unsafe behaviour of agents in stochastic environments. With that, we can, for instance, enforce safety in reinforcement learning. Supplementary material, such as a demonstrator on a two-player version of the classic computer game snake can be found here.
Our paper Adaptive Shielding under Uncertainty has been accepted for publication at ACC 2021.
You can already check out the supplementary material of the paper at https://adaptiveshielding.xyz/. We use probabilistic verification, model refinement, and online estimation of transition probabilities to adaptively create shields that improve existing controllers.
I have recently joined the Schaffhausen Institute of Technology (SIT) as a postdoctoral researcher, where I work in the team of Prof. Mauro Pezzè.
While the current global situation is difficult, I am looking forward to exciting research opportunities in a new and fascinating environment in the beautiful city of Schaffhausen, Switzerland.
Yesterday, I successfully defended my doctoral thesis with the title “Learning-Based Testing in Networked Environments in the Presence of Timed and Stochastic Behaviour”.
You can find my thesis under my publications and here.
We have received the best paper award at ICTSS 2019 for our paper “Learning a Behavior Model of Hybrid Systems Through Combining Model-Based Testing and Machine Learning”.
The work presented in this paper is the result of a collaboration of nine researchers from four institutes at Graz University of Technology.
Our survey on ” model-based testing and model learning” got published. In this survey, we review combinations of learning and testing, which can roughly be classified into two categories:
- Test-based learning: techniques in this category apply testing techniques to enable learning models of software systems.
- Learning-based testing: techniques in this category generally learn models in order to use them as a basis for model-based testing.
The survey is part of a book on the Dagstuhl seminar “Machine Learning for Dynamic Software Analysis: Potentials and Limits” and can be found here and in my publications.