Probabilistic Safety Guarantee for Stochastic Control Systems Using Average Reward MDPs

The recent publication of a novel algorithm for stochastic control systems addresses the critical challenge of ensuring safety amidst random noise. By reducing the safety objective to an average reward Markov Decision Process (MDP), the algorithm enables the computation of safe policies that maintain high confidence levels throughout the uncertain evolution of state variables. This advancement is particularly relevant for systems like the Double Integrator and Inverted Pendulum, where traditional methods often fall short. Numerical validation demonstrates that the average-reward MDP solution not only converges faster but also provides higher quality outcomes compared to the minimum discounted-reward solution. This development is significant as it enhances the reliability of control systems in unpredictable environments, paving the way for safer applications in various fields.