Showing 1–1 of 1 results for author: Dehnert, C

Safety-Constrained Reinforcement Learning for MDPs

Authors: Sebastian Junges, Nils Jansen, Christian Dehnert, Ufuk Topcu, Joost-Pieter Katoen

Abstract: We consider controller synthesis for stochastic and partially unknown environments in which safety is essential. Specifically, we abstract the problem as a Markov decision process in which the expected performance is measured using a cost function that is unknown prior to run-time exploration of the state space. Standard learning approaches synthesize cost-optimal strategies without guaranteeing s… ▽ More We consider controller synthesis for stochastic and partially unknown environments in which safety is essential. Specifically, we abstract the problem as a Markov decision process in which the expected performance is measured using a cost function that is unknown prior to run-time exploration of the state space. Standard learning approaches synthesize cost-optimal strategies without guaranteeing safety properties. To remedy this, we first compute safe, permissive strategies. Then, exploration is constrained to these strategies and thereby meets the imposed safety requirements. Exploiting an iterative learning procedure, the resulting policy is safety-constrained and optimal. We show correctness and completeness of the method and discuss the use of several heuristics to increase its scalability. Finally, we demonstrate the applicability by means of a prototype implementation. △ Less

Submitted 20 October, 2015; originally announced October 2015.