Synchronization in Markov decision processes

Markov Decision Processes (MDPs) are 1-1/2 player stochastic games. The player tries to maximize the probability to satisfy an objective. Traditionally, the objective of the player is expressed as a set of desired sequences of states visited during the game. Recently, MDPs are viewed as generators of probability distributions over states, and objectives are defined as sets of sequences of probability distributions. We study synchronizing objectives that require that some state tend to accumulate all the probability mass. We consider three winning modes: sure, almost-sure and limit-sure.

Authors

• Mahsa Shirmohammadi
• Laurent Doyen
• Thierry Massart

Keywords

No keywords are indexed for this paper.