| Summary: | Designing systems with the ability to make optimal decisions under uncertainty is one of the goals of artificial intelligence. However, in many applications the design of optimal planners is complicated due to imprecise inputs and uncertain outputs resulting from stochastic dynamics. Partially Observable Markov Decision Processes (POMDPs) provide a rich mathematical framework to model these kinds of problems. However, the high computational demand of solution methods for POMDPs is a drawback for applying them in practice. === In this thesis, we present a two-fold approach for improving the tractability of POMDP planning. First, we focus on designing good heuristics for POMDP approximation algorithms. We aim to scale up the efficiency of a class of POMDP approximations called point-based planning methods by designing a good planning space. We study the effect of three properties of reachable belief state points that may influence the performance of point-based approximation methods. Second, we investigate approaches to designing good controllers using an alternative representation of systems with partial observability called Predictive State Representation (PSR). This part of the thesis advocates the usefulness and practicality of PSRs in planning under uncertainty. We also attempt to move some useful characteristics of the PSR model, which has a predictive view of the world, to the POMDP model, which has a probabilistic view of the hidden states of the world. We propose a planning algorithm motivated by the connections between the two models.
|
|---|
