Formal models of cyber-physical systems, such as priced timed Markov decision processes, require a state space with continuous and discrete components. The problem of controller synthesis for such systems then can be cast as finding optimal strategies for Markov decision processes over a Euclidean state space. We develop two different reinforcement learning strategies that tackle the problem of continuous state spaces via online partition refinement techniques. We provide theoretical insights into the convergence of partition refinement schemes. Our techniques are implemented in Open image in new window . Experimental results show the advantages of our new techniques over previous optimization algorithms of Open image in new window .
|Konference||International Symposium on Automated Technology for Verification and Analysis|
|Periode||28/10/2019 → 31/10/2019|
|Navn||Lecture Notes in Computer Science|