Fifth edition AI lunch meeting (TU/e internal)

image

Event Details

Date
Friday September 14, 2018 from 12:00 PM to 1:30 PM
Location
TU/e, Gemini Zuid - 4.24
Organizer
DSC/e, HTSC & Center H&T

Many TU/e researchers are advancing or using Artificial Intelligence in their work. However there is currently no overview, no coordination and no strengthening of individual activities. For this reason we joined forces and organize a series of (internal) lunch meetings.

Program

Start

Speaker

Title

12:00

Johan Lukkien (Dean M&CS)

Introduction

12:15

Yingqian Zhang, Information Systems (IE&IS)

Fair allocation in multi-agent systems

12:35

Roland Toth, Control Systems (EE)                 

Machine Learning Based Identification of Dynamical Systems

12:55

Rui Castro, Statistics (M&CS)

Learning to learn: closing the loop between data analysis and acquisition

13:15

 

Wrap up

ABSTRACTS

Yingqian Zhang
Traditional task allocation problems have focused on achieving one shot optimality, which typically aims at finding the minimum cost allocations.
In practice, many task or resource allocations are of repeated nature, where the allocation outcome of previous rounds may influence the players’ decision to participate in subsequent rounds. We consider a fair task allocation problem in transportation with multiple agents (players), and propose an efficient algorithm for matching tasks in a fair way among players.
We then demonstrate by simulations that the proposed algorithm incentivizes players to keep participating. With more participants, more resources are available in the system, which eventually lead to a higher social welfare.
In many cases such as applications in circular economy, such a fair matching algorithm is more socially desired than the classical cost minimization ones.

Roland Toth
Rising performance expectations have been a driving force in engineering to consider modelling and control of dynamical systems with increasingly nonlinear, time-varying and spatial behaviour. Next to this, the parallel increase of complexity of these engineering systems has resulted in a growing need to develop data-driven methods that can provide fast and reliable models of diverse dynamical phenomena without the cumbersome process of first principles based modelling. In this talk, it will be shown how successful methods from Machine Learning, like Gaussian Processes, Support Vector Machines, and Deep Learning can provide efficient tools to meet the growing challenges in identification of dynamical systems.

Rui Castro
In many scenarios where machine learning and statistics tools are used one typically starts with a set of previously collected data, in order to learn generic relations between the various observed variables.  This means such approaches rely on passive data collection methodologies.  However, in many practical situations it is possible to adjust the data collection process based on information gleaned from previous observations, in the spirit of the "twenty-questions" game, in what is known as adaptive sensing, active learning, or sequential design of experiments.  The intricate relation between data analysis and acquisition in adaptive sensing paradigms is extremely powerful, but creates complicated data dependencies making it challenging to devise simultaneously practical and theoretically sound learning methodologies.  In this talk I focus primarily on estimation and detection of high-dimensional sparse signals in noise, for which it is possible to devise (near) optimal adaptive sensing procedures that provably outperform the best possible inference methods based on non-adaptive sensing.

image
Organizer

Data Science

Data Science is an interdisciplinary field that uses a variety of techniques to create value based on extracting knowledge and insights from available data. The successful and responsible application of these methods highly depends on a good understanding of the application domain, taking into account ethics, business models, and human behavior.