...
 
Commits (2)
\documentclass{article}
\usepackage{geometry}
\usepackage[utf8]{inputenc}
\usepackage{color}
\usepackage{xcolor}
\usepackage{biblatex} %[style=alphabetic,citestyle=authoryear] citestyle=numeric-comp
\usepackage{longtable}
\usepackage{graphicx}
\usepackage{subcaption}
% Comment out the line below before submitting, so that any remaining to-do:s become errors instead of showing up in the proposal.
\newcommand{\TODO}[1]{\textcolor{red}{[\textbf{TODO:} \textit{#1}]}}
% convenience function for citing consortium's research in section 5
\newcommand{\mycite}[1]{\cite{#1}:\\ \AtNextCite{\defcounter{maxnames}{99}}\fullcite{#1}}
% A macro for marking text that may or may not fit in the draft proposal due to the page-count limit.
% Text marked \priority[1]{...} will show up as usual, while text marked \priority[2]{...} (or higher) will show up in grey, and can be made to disappear entirely.
\newcommand{\priority}[2][1]{%
\ifnum#1>1% the number at the end of this line is the lowest priority that will still show up normally
\textcolor{gray}{#2}% comment out this line to remove priority-2 text entirely.
\else%
#2%
\fi%
}
\title{Research Training Group on \\ Learning, Estimation, Control, and Optimisation (LECO)}
\author{ }
\date{May 2019}
......@@ -152,7 +166,7 @@ Examples of thematic areas suitable for PhD projects include:
\begin{itemize}
\item To investigate how latent spaces can be shaped efficiently to learn suitable representations for the optimal control of forward dynamic systems. The motivation is that neural networks are powerful function approximators but lead to complex latent state representations with poor generalisation capacities when used out of the box. \TODO{Clarify where the link between machine learning and optimal control lies.}
\item To combine model-free reinforcement learning with model-based real-time control of dynamical systems. This has several applications and motivations. On the one hand, learning techniques can be used for efficient processing of high-dimensional real-world observations (e.g. video and audio data) and for learning the cost function in model predictive control. On the other hand, model predictive control can be used to exploit convexity, provide safety bounds, and generate artificial training data. Two specific real-world applications targeted by the involved research groups are high-level arm control in robotics and safe control of racing cars. A particular focus will be on efficient and robust numerical implementations of these problems.
\priority[2]{\item To combine model-free reinforcement learning with model-based real-time control of dynamical systems. This has several applications and motivations. On the one hand, learning techniques can be used for efficient processing of high-dimensional real-world observations (e.g. video and audio data) and for learning the cost function in model predictive control. On the other hand, model predictive control can be used to exploit convexity, provide safety bounds, and generate artificial training data. Two specific real-world applications targeted by the involved research groups are high-level arm control in robotics and safe control of racing cars. A particular focus will be on efficient and robust numerical implementations of these problems.}
\item To investigate the aspect of safety in real-world controlled systems such as racing cars, robotic arms, or wind turbines. Here we aim to combine rigorous, provably safe but often conservative techniques from robust model predictive control with flexible and scalable machine learning technologies. \TODO{cite work by Zeilinger et al.}
......@@ -231,6 +245,50 @@ We will investigate data-driven methods for improving efficient learning and opt
% - Outline of the qualification programme based on the research programme
% - Short description of the supervision strategy
\TODO{A text describing the qualification process, emphasising objectivity, gender equality, transparency\ldots}
\TODO{Points to make about the supervision strategy (from docs)}
%Exchange “guest lectures” in regular courses.
%Yearly GK Meeting
%Supervision strategy: by two professors and one “postdoc mentor” (usually from one of the two teams)
%Each PhD needs:
% Advisory board
% Progress report
%Yearly meeting with advisors, e.g. at academies or summer/winter/spring schools
%Process to ensure progress of work
%Further program ideas:
% Tag der offenen Tür / Exhibit Day
% GK internal events:
% Sommer/Winter-Akademien mit Profs (+ Externe)
% Anregungen
% Visionsbildung
% Evaluierung
% Progress reports
% Retreats der PhDs
% Vorstellung der eig. Projekte
% Austausch über Erfahrungen und Arbeitsweisen
% Reading groups
% Seminarreihe für die PhDs
% Lab-rotations (der PhDs)
% Internal Reviews - For co-authorship
% Interdisciplinary co-supervision
% Shared workspace to enhance exchange and
% Gemeinsamer arbeitsraum zur Förderung des Austauschs und um Gemeinschaftsgefühl zu stärken (Idee, Tobi)
% External stay (secondments) at industry partners, (to attract more qualified applicants, contributes to the environment)
% Provide tailored coding and research infrastructure to speed up and simplify research (in our case mostly software). Can be collaborative, e.g. some PhDs work on framework while others apply the framework
% Soft-skill courses:
% Good scientific practice
% Scientific writing (paper and application)
% Ethical aspects
% Gender!
% Objective: Wissenschaftliche Selbstständigkeit der PhDs (founded on Student initiative)
% Idea: PhDs don’t start all at the same time ⅓ right away, ⅓ 0.75 years later, ⅓ 1.5 years later (to ensure continuity and build up supervision competence)
\priority[2]{
Existing and planned (*) graduate level lectures and seminars relevant for LECO doctoral students:
\begin{enumerate}
\item
......@@ -279,6 +337,7 @@ L - Machine Learning (incl. Reinforcement Learning)
\item Robotics \TODO{Bosch}
\end{enumerate}
\end{enumerate}
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Environment}
......@@ -292,10 +351,29 @@ L - Machine Learning (incl. Reinforcement Learning)
% - Description of why the proposed location is suitable for the topic of this Research Training Group
% - Statement by the university on how the Research Training Group fits into the university’s environment and what structural innovations are anticipated.
The University of Freiburg provides an excellent scientific environment, with over 24 000 students and over 4 000 researchers it is the largest employer in Freiburg. The city is known for its environmentally friendly policies making it an attractive place to live and work, which attracts also international researchers.
The technical faculty is renowned for it's outstanding research in
Specific to this research programme,
The University of Freiburg provides an excellent scientific environment, with over 24 000 students from over 100 nations and over 4 500 researchers it is the largest employer in Freiburg. The city is known for its environmentally friendly policies and is situated in a diverse landscape making it an attractive place to live and work. This setting also attracts international researchers.
With nearly 250 different study programs offered at 11 faculties the University of Freiburg provides also a diverse environment which creates an ideal breeding ground for interactions and collaborations between different research fields. For this Research Training Group, LECO, researchers from three different faculties cooperate to form a interdisciplinary consortium that bridges gaps between statistics/stochastic/mathematical foundations, systems and control theory, and machine learning.
Even thought these fields are strongly related and share common foundation, there has been little interaction between the respective research communities leading to many re-discoveries of algorithms and methods already known to another community.
To overcome this problem and to achieve mutual understanding and collaborative research, leading researchers from these three fields have gathered to form LECO Research Training Group.
The aim is to train students that are familiar with the methods, way of thinking, and language of all of these research fields, making them interpreters that can translate between the different communities. These students will have access to a huge pool of knowledge, that enables them to detect and exploit synergies between the fields and aids fast progress of research that is likely to lead to substantial contributions.
Novel methods will be applied in finance, bioinformatics, and robotics, which makes them available to a broad research community and ensures general applicability.
The University of Freiburg is part of Eucor, a unique trinational university alliance, that represents a diverse scientific environment that bridges boarders. This unified campus will be available to the LECO fellows, giving them to access the study program of all Eucor partners and setting the ground work for research powered by the combined resources of all 5 Universities.
%
\begin{figure}
\centering
\includegraphics[width=0.32\linewidth]{figures/robotics_freiburg_stats.png}
\hfill
\includegraphics[width=0.32\linewidth,trim=10pt 0pt 0pt 0pt, clip]{figures/ml_freiburg_stats.png}
\hfill
\includegraphics[width=0.32\linewidth]{figures/ai_freiburg_stats.png}
\caption{Citations of University of Freiburg in Robotics, Machine Learning, and Artificial Intelligence in comparison to other German research institutions.}
\label{fig:citation_stats}
\end{figure}
%
The technical faculty is renowned for it's outstanding research in Robotics, Machine Learning, and Artificial Intelligence. As of 17.02.2019 the University of Freiburg was the leading research institute in Germany in all of these three fields. A comparison to other research institutes in Germany is shown in Figure~\ref{fig:citation_stats}.
\subsection{External Academic Contacts and Planned Cooperation Topics}
\begin{itemize}
......