|Machine Learning for Autonomous Robots
Martin Riedmiller (University of Osnabrueck) |
In recent years, many successful applications of machine learning
methods have been developped, e.g. in classification, diagnosis or
forecasting tasks. However, for a broad acceptance of learning
methods as a standard software tool, still many theoretical and
practical problems have to be solved. This is especially true for
Reinforcement Learning scenarios, where the only training signal is
given in terms of success or failure. Although this paradigm is in
principle very powerful due to the minimal requirements on training
information, today real world applications often fail due to the large
amount on training experiences, until a task is successfully learned.
Our research effort lies in narrowing this gap by developing methods
for data-efficient and robust machine learning. Autonomous robots -
some of them with the ability to play soccer - are one of our
favorite testbeds. Examples of efficient machine learning methods and
their integration into large software systems are shown on several
(real world) tasks.
Prof. Dr. Martin Riedmiller is chair of the neuroinformatics group at
the University of Osnabrueck, Germany. His research interests
are machine learning and robotics. An important goal of his research
activities is to make machine learning algorithms successfully
work in practical applications. Several industrial projects resulted
from this, e.g. forecasting systems for financial markets
and newspapers sales rates. One of his main current research focuses is
the investigation of reinforcement learning algorithms
for complex real world applications, especially for autonomously
learning robots. His robotic soccer team 'Brainstormers',
a hybrid approach of machine learning, AI and conventional program code,
was always among the best three teams during last
years international competitions.
|Modelling and Querying Interaction Networks in the Biochemical Abstract Machine BIOCHAM
|François Fages (INRIA Rocquencourt)|
In recent years, molecular biology has engaged in a large-scale effort
to elucidate high-level cellular processes in terms of their biochemical basis
at the molecular level. The mass production of post genomic data,
such as ARN expression, protein production and protein-protein interaction,
raises the need of a strong parallel effort on the formal representation
of biological processes.
We shall present the Biochemical Abstract Machine BIOCHAM
and advocate its use as a formal modeling environment for networks biology.
Biocham provides a precise semantics to biomolecular interaction maps.
Based on this formal semantics, the Biocham system offers automated reasoning tools
for querying the temporal properties of the system under all its possible behaviors.
We shall review the main features of Biocham and report on our modeling experience
with this language. In particular we shall report on a model of the mammalian cell cycle's
control developped after Kohn's map.|
Dr. Hab. François Fages leads the Contraintes group at INRIA Rocquencourt.
His research is focused on Constraint Programming languages,
from their fine relationship with mathematical logic (including linear logic),
to new proof methods, type systems and new execution models
(combining constraint propagation with state change), and on two application domains:
combinatorial optimization problems and Systems Biology.
Before his recruitment at INRIA in 1999, Francois Fages was a permanent researcher at CNRS
Ecole Normale Supérieure in Paris since 1983, the date of his Doctorate Thesis.
His main research interests have been unification theory,
automated deduction, rule-based reactive programming and non-monotonic reasoning,
constraint logic programming, concurrent constraint programming and systems biology.
|Impact and Perspective of Planning Techniques in Robotics
Malik Ghallab (LAAS-CNRS) |
techniques can contribute to robotics by making it easier to program a
robot and by augmenting its autonomy, usefulness and robustness.
Planning is certainly needed if a robot has to face a wide diversity of
tasks and/or a variety of environments. When planning is integrated
within a robot, it usually takes several forms and is implemented
throughout different systems, often relying on domain-specific
representations and techniques. Among these various forms of robot
planning, there is in particular path and motion planning, perception,
planning navigation planning, manipulation planning, and domain
independent task and mission planning. This talk illustrates some of
these techniques and discusses their impact and perspective for the
development of Robotics.
Prof. Ghallab is director of the LAAS-CNRS
institute in Toulouse. His research interests in robotics and AI are in
the integration of perception, action and reasoning capabilities within
autonomous robots. His activity is mainly focused on mobile robotics,
within experimental projects and research areas such as perception,
control, planning and decision making. He contributed to topics such as
object recognition, scene interpretation, heuristics search, pattern
matching and unification algorithms, knowledge compiling, temporal
planning and supervision systems.
|Computer supported Mathematics
Jörg Siekmann (DFKI, University of Saarbrücken)
The year 2004 marks the fiftieth birthday of the
first mathematical theorem to be proved by a computer: “the sum of two
even numbers is again an even number” (with Martin Davis’
implementation of Presburger Arithmetic)
Classical theorem proving procedures of today are based on ingenious
search techniques to find a proof for a given theorem in very large
search spaces – often in the range of several billion clauses.
The shift from search based methods to more abstract planning
techniques opened up a new paradigm for mathematical reasoning on a
computer and several systems of the new kind employ a mix of classical
as well as proof planning techniques.
In my talk I shall trace some key ideas of the past and then
concentrate on current systems based on proof planning, using ΩMEGA, a
mathematical assistant system and ActiveMath, a mathematical learning
environment as case studies for demonstrating current strengths (and
weaknesses) of the field.
Siekmann is a scientific director of the German Research Centre for
Artificial Intelligence (DFKI) at Saarbrücken and one of the
founders of A.I. in Germany.
His research interest in automated reasoning is mainly focused on proof
presentation, graph based theorem proving, knowledge representation for
mathematics and proof planning as well as unification theory.
As a scientific director of the DFKI and a professor of computer
science at the Universität des Saarlandes he is the head of four
research labs: multi agent systems, safety and security verification,
automated reasoning and e-learning, each encompassing about a dozen
researchers enjoying the interdisciplinary nature and cross
fertilization of research on the borderline between these areas.
Wolfgang Wahlster (DFKI, University of Saarbrücken)|
Fortschritte auf dem Gebiet der mobilen
Breitbandkommunikation und des semantischen Web ermöglichen
innovative Internetdienste, die sich durch Personalisierung und
Lokalisierung auszeichnen. Die ontologische Beschreibung von
Webdiensten sowie die Informationsextraktion mithilfe der
Sprachtechnologie ermöglicht den Übergang vom Suchen zum
gezielten Finden von Information und Diensten im Web. Der Vortrag
illustriert die neuartigen Informatik-Methoden anhand praktischer
Beispiele aus dem Bereich UMTS-Mehrwehrtdienste, Ambient Intelligence
und Infotainment im Automobil.
Prof. Wahlster ist der Leiter des Deutschen Forschungszentrums für
Künstliche Intelligenz und Professor für Informatik in
Saarbrücken. 2001 wurde er mit dem Zukunftspreis des
Bundespräsidenten ausgezeichnet. 2003 wurde er in die Nobelpreisakademie
in Stockholm aufgenommen. Er leitet die ISTAG-Arbeitsgruppe zur
Definition der Förderthemen für das siebte EU-Rahmenprogramm.