Decision Making
In November 2020, Jun.-Prof. Dr. Setareh Maghsudi has been appointed as W2-Professor for "Decision Making" at the Department of Computer Science of the University of Tübingen. She is being supported with Cluster funding for a research project.
Today’s world is characterized by the massive growth of information-thirsty human-driven devices as well as smart machines that act without human interventions. In a dynamic, stochastic, and communication-constrained environment, while being uncertain about the actions of their counterparts, the devices and machines engage in real-time interaction to fulfill some specific goals that might constitute conflict or cooperation. Thereby, they create large dynamic networks and produce excessive amounts of data that result from observing side-information and possibly-limited feedbacks of actions.
We concentrate on the analysis and optimization of such systems by developing efficient and convergent decision-making strategies. From the theoretical perspective, the developed analytical methods lie at the intersection of game theory, artificial intelligence, and data science. The proposed methods are forward-looking and find application not only in conventional research fields such as resource allocation in wireless communication networks, but also in several innovative directions, including the Internet of Things and digital platforms.
| Distributed Machine Learning over Unreliable Networks |
Cluster funded research project
- Team member: Behzad Nourani Koliji (PhD since February 2021)
Online learning is a variety of sequential decision-making problems under uncertainty. Often, the challenge is envisioned by an agent interacting with an unknown environment. The agent successively selects an action from a set of available actions. The agent obtains some a priori-unknown and state-dependent reward upon choosing each action. A solution to this problem is a decision-making policy to select an action at every round. Often, the evaluation follows in terms of regret bound or discounted accumulated reward.
Graph signal processing (GSP) is a recently-established branch of signal processing. GSP methods handle the excessive amount of data that is collected in various circumstances in a fast and efficient manner.
This project establishes a link between GSP and online decision-making. It develops, adapts, and integrates the GSP tools to enable fast and efficient online learning and decision-making under uncertainty. For example, the graph nodes can represent actions, while each edge shows the relation between the reward generating processes of a pair of actions. The instantaneous rewards of each action then form a time-series at each node of the graph. In this model, relationships such as causality or correlation are inferable by taking advantage of GSP methods. Such information improves the learning rate and reduces the complexity significantly, as they allow to take advantage of mutual information or the structure for efficient sampling. The challenges include dealing with dynamic and time-varying graphs, addressing uncertainties in the graph structure, and managing ultra-large/dense graphs.
Contact
Prof. Dr.-Ing. Setareh Maghsudi
Decision Making
University of Tübingen
Department of Computer Science
Maria-von-Linden-Straße 6, 2nd floor
Room No. 20-5/A19
720740 Tübingen
setareh.maghsudi@uni-tuebingen.de