Cost-Optimal Control of Markov Decision Processes Under Signal Temporal Logic Constraints

We present a method to find a cost-optimal policy for a given finite-horizon Markov decision process (MDP) with unknown transition probability, such that the probability of satisfying a given signal temporal logic specification is above a desired threshold. We propose an augmentation of the MDP state space to enable the expression of the STL objective as a reachability objective. In this augmented space, the optimal policy problem is re-formulated as a finite-horizon constrained Markov decision process (CMDP). We then develop a model-free reinforcement learning (RL) scheme to provide an approximately optimal policy for any general finite horizon CMDP problem. This scheme can make use of any off-the-shelf model-free RL algorithm and considers the general space of non-stationary randomized policies. Finally, we illustrate the applicability of our RL-based approach through two case studies.