Roderick BloemTU Graz, Austria
Shield Synthesis for Safe Reinforcement Learning
You have a reinforcement learning system? Sure, it works great, but does it give you any guarantees? I thought not. We will describe methods to use reactive synthesis to construct runtime enforcement modules (shields) that can ensure that a system works correctly, even if the system has bugs. If the system doesn't have too many bugs, the behavior of the shielded system will stay close to the behavior of the original system. We will show extensions to probabilistic and timed settings.
Anca MuschollLaBRI, Bordeaux, France
Deterministic, sound negotiations represent a tractable distributed model related to Petri nets and Zielonka automata. We describe in this talk how to exploit the structure of sound negotiations in order to derive polynomial-time algorithms, avoiding the state explosion induced by configurations. We also explain how to obtain polynomial-time Angluin-style algorithms for active learning of sound negotiations. (This is joint work with I. Walukiewicz.)
Nicola OlivettiAix-Marseille University, France
Non-normal modal logics vindicated
Non-normal modal logics (NNML) are known since a long time. They are a generalisation of standard normal modal logics, that do not satisfy some basic principles of them, in some cases they even contradict them. NNML have been studied mainly by philosophers for overcoming difficulties and paradoxes of epistemic and deontic reasoning originated by their formalisation in normal modal logics. In my opinion the interest of NNML has not been fully recognised in Knowledge Representation, not only for epistemic and deontic reasoning, but also for reasoning about agency and capability. One possible reason that hindered their larger use is the rather abstract Neighbourhood semantics which comes with them, as well as the lack of efficient and "friendly" proof systems. The aim of my talk is to vindicate Non Normal modal logics, suggesting also a natural reformulation of their semantics and recent developments in their proof theory.
S.E. (Sicco) VerwerTU Delft, The Netherlands
State machine learning in practice using Flexfringe: flexibility, use-cases, and interpretability
State machine learning deals with the problem of reconstructing a state machine (automaton) model from sequences of observations. In the standard use-case, we have access to the input and output events from a software component and the goal is to find a small model describing the mapping from inputs to outputs. In more complex settings, we observe only inputs , only have access to indirect signals from e.g. sensors, or sometimes we may have white-box access to the system. In this talk, I will show examples of standard and complex settings and how to setup Flexfringe to learn models in each of these cases. I will also demonstrate how the obtained models can be used in further processes such as finding bugs and safety verification. A special focus is on the trade-off between predictive accuracy and interpretability of these models. Reverse engineering models of systems is often performed for interpretability while the learning algorithms focus on accuracy. I will present initial ideas on how to modify this learning objective and results we obtained using it.