Fast Learning of Small Strategies

Jan Křetínský
TU Munich
Friday, 30 March, 2018 - 14:00
ULB La Plaine NO, Salle Solvay
In verification, precise analysis is required, but the algorithms usually suffer from scalability issues. In machine learning, scalability is achieved, but with only very weak guarantees. We show how to merge the two philosophies and profit from both. In this talk, we discuss models such as Markov decision processes, two-player games, stochastic games, and objectives such as reachability, LTL, mean payoff. We show how to learn ε-optimal strategies fast and how to represent them concisely so that some understanding of the behaviour and debugging information can be extracted.