Tobias Meggendorfer

(Department of Computer Science, RPTU)
hosted by Department of Computer Science

"Quantifying Risk in Decision Making"

A central (or rather: the ultimate) problem in artificial intelligence is to choose which action to take in any given situation. The standard route to approach this question is to define what it means for an action to be good or bad. For example, taking an action that yields money could be labelled as "good". Now, once probabilities come into play, the question arises how we judge, e.g., a lottery - choosing to play a lottery does not yield one fixed amount of money, but rather the money we obtain is drawn from a distribution. Typically, the focus of (automated) decision making lies exclusively on the expectation of this lottery. However, consider a lottery that with a one-in-million chance yields two million times your wager: Here, the expectation-optimal strategy is to bet everything you have - after all, it gets doubled in expectation! Yet, in many situations we would like to not take that risk. In this talk, I will illustrate some initial work that tries to tackle this question by quantifying the risk associated with a course of actions in probabilistic systems such as Markov decision processes.

Time:	Friday, 12.01.2024, 13:00
Place:

Termin als Datei downloaden und in den Kalender importieren.