Guest Talk: Raphaël Reynouard: The Baum-Welch Algorithm in All its Power

Tuesday, March 29, 2022, 11:00am

Location: RWTH Aachen University, Department of Computer Science - Ahornstr. 55, building E3, room 9u10

Speaker: Raphaël Reynouard



Cyber-physical systems (CPSs) are naturally modelled as reactive systems with nondeterministic and probabilistic dynamics. Model-based verification techniques have proved effective in the deployment of safety-critical CPSs.

The construction of an accurate formal model for the system is central to the successful application of such techniques. Manual construction can be a resource-demanding and error-prone process, thus motivating the design of automata learning algorithms to synthesise a system model from observed system behaviours.

In this presentation, we revisit and adapt the classic Baum-Welch algorithm for learning several kinds of Markov models, namely: Hidden Markov Models, Labeled Makov Chains, Markov Decision Processes, Continuous Hidden Markov Models, Continuous-Time Markov Chains.

For the case of MDPs, which typically demand more observations, we present a model-based active learning sampling strategy that chooses examples that are most informative w.r.t. the current model hypothesis.