November 7, 2016Nicolas Basset (ULB)
We focus on stochastic games,
which can model interaction with an adverse environment,
as well as probabilistic behaviour arising from uncertainties.
Our contribution is twofold.
First, we study long-run specifications
expressed as quantitative multi-dimensional
mean-payoff and ratio objectives.
We then develop an algorithm to synthesise epsilon-optimal strategies for conjunctions of almost sure satisfaction for mean payoffs
and ratio rewards (in general games) and Boolean combinations of expected
mean-payoffs (in controllable multi-chain games).
Second, we propose a compositional framework, together with assume-guarantee rules,
which enables winning strategies synthesised for individual components to be composed to a winning strategy for the composed game.
The framework applies to a broad class of properties, which also include expected total rewards,
and has been implemented in the software tool PRISM-games.