Povzetek:

Partially observable Markov decision processes (POMDPs) have been known for approximately fifty years and are still popular today. A Markov decision process (MDP) takes into account the stochastic uncertainty in a system's action-effects. MDPs also consider the value of actions in different states of a system. The aim is then to develop algorithms to effectively and efficiently find actions to maximize the values that actions bring to a system in the long-run. POMDPs generalize MDPs by also dealing with noise in system sensors.

In this course, you will learn what an MDP is and how it is used for planning in a fully-observable environment. Then you will learn what a POMDP is and a basic algorithm for online planning in partially observable environments. Finally, you will see an example of online POMDP planning used in an agent architecture which combines techniques from the Belief-Desire-Intention formalism for sophisticated goal management. The whole course will take three hours: MDPs = 1 hour, POMDPs = 1 hour, POMDP integration into BDI architecture = 1 hour.

O predavatelju:

Gavin Rens is a post-doctorate researcher in the department of computer science at the University of Cape Town, South Africa. Before this, he held a post-doctorate position at the University of KwaZulu-Natal, South Africa, where he also received his Ph.D, in 2014. Gavin was awarded a studentship through the German student exchange service (DAAD) to perform one year of his doctoral research at the RWTH Aachen University in Germany. He received his BSc, Honours BSc and MSc (the latter with distinction), all from the University of South Africa (UNISA). Gavin has published twenty papers on a range of topics in the area of knowledge representation, reasoning and decision-making under uncertainty.