Reachability in MDPs: Refining Convergence of Value Iteration
Markov Decision Processes (MDP) are a widely used model including both non-deterministic and probabilistic choices. Minimal and maximal probabilities to reach a target set of states, with respect to a policy resolving non-determinism, may be computed by several methods including value iteration. This algorithm, easy to implement and efficient in terms of space complexity, consists in iteratively finding the probabilities of paths of increasing length. However, it raises three issues: (1) defining a stopping criterion ensuring a bound on the approximation, (2) analyzing the rate of convergence, and (3) specifying an additional procedure to obtain the exact values once a sufficient number of iterations has been performed. The first two issues are still open and for the third one a “crude” upper bound on the number of iterations has been proposed. Based on a graph analysis and transformation of MDPs, we address these problems. First we introduce an interval iteration algorithm, for which the stopping criterion is straightforward. Then we exhibit convergence rate. Finally we significantly improve the bound on the number of iterations required to get the exact values.
Unable to display preview. Download preview PDF.
- 1.Baier, C., Katoen, J.-P.: Principles of Model Checking. MIT Press (2008)Google Scholar
- 2.Brázdil, T., Chatterjee, K., Chmelík, M., Forejt, V., Křetínský, J., Kwiatkowska, M., Parker, D., Ujma, M.: Verification of Markov decision processes using learning algorithms. Research Report arXiv:1402.2967 (2014)Google Scholar
- 4.de Alfaro, L.: Formal Verification of Probabilistic Systems. PhD thesis, Stanford University (1997)Google Scholar
- 6.Haddad, S., Monmege, B.: Reachability in MDPs: Refining convergence of value iteration. Technical Report LSV-14-07, LSV, ENS Cachan (2014), http://www.lsv.ens-cachan.fr/Publis/RAPPORTS_LSV/PDF/rr-lsv-2014-07.pdf
- 7.Katoen, J.-P., Zapreev, I.S.: Safe on-the-fly steady-state detection for time-bounded reachability. In: QEST 2006, pp. 301–310 (2006)Google Scholar