Springe direkt zu Inhalt

Disputation Andreas Philipp

02.12.2021 | 09:00
Thema der Dissertation:
Perception and Prediction of Urban Traffic Scenarios for Autonomous Driving
Thema der Disputation:
Partial Observable Markov Decision Process (POMDP)
Abstract: Markov Decision Processes provide a framework for solving sequential decision problems under uncertainty. The uncertainty concerns the outcomes of the selected actions and in case of partial observability, also the current state of the agent. The Markov assumption states that the optimal decision depends only on the current state of the world and is independent of the previous history. There are various methods of finding an optimal policy, depending on the nature of uncertainty, the environment and the action space of an agent.
This presentation explains the Markov assumption and gives an overview of various Markov models based on this assumption. It presents the basic Markov Decision Process (MDP) using the 4x3 environment example of Russel & Norvig. The main topic is the Partial Observable Markov Decision Process (POMDP), which generalizes the MDP. A proven solution method is presented using the finite state environment example of Thrun, Burgard and Fox. Thereafter, a real-time capable online solution method, presented in a recent research paper, is discussed. Finally, the major benefits and problems of the approaches are summarized.

Zeit & Ort

02.12.2021 | 09:00