On some extensions of discounted Markov decision processes
- A+
：Yi Zhang (University of Birmingham)
： 2024-10-20 09:30
：Conference Room C610 at Administration Building at Haiyun Campus
Speaker：Yi Zhang (University of Birmingham)
Time：2024-10-20, 9:30
Location：Conference Room C610 at Administration Building at Haiyun Campus
Abstract:
Discounted Markov decision processes have been studiedintensively in the literature. In this talk, I would like to present two extensions of the discounted Markov decision processes (MDP), and report some results for the more general models. The first extension comes from the fact that discounted MDPs can be viewed as special uniformly absorbing MDPs, which in turn are special models of absorbing MDPs. I like to present optimality results for absorbing MDPs and discuss some results that are valid for uniformly absorbing MDPs but not for the more general models. The second extension comes from the fact that in the standard discounted MDP, the discount function gamma is linear: if present value of the utility x in the next stage is given by gamma(x)=beta x. We may extend this linear discount function to a more general discount function. We observe that under certain conditions, this extension can be reduced to an equivalent two-player stochastic game model with perfect information. Some results presented in this talk are from joint works with Alexey Piunovskiy, Ernst Presman and Xinran Zheng.
