Conceptual

Data-Driven Interval MDP Abstraction for Formal Control of Stochastic Nonlinear Systems

A technique that abstracts a discrete-time nonlinear system with additive stochastic noise into a finite-state interval Markov decision process using only sample access to the dynamics and noise. Backward reachable sets are underapproximated from forward simulations, and transition-probability intervals are Clopper-Pearson confidence intervals, so each interval carries a probably-approximately-correct guarantee. Policies synthesized on the abstraction satisfy a reach-avoid objective on the concrete system with PAC guarantees despite unknown dynamics.