Projects

Here is a collection of some of my favourite former and current projects - feel free to explore and exploit ...

(2015 - present)

Data-efficiency is a core requirement for artificial intelligence. We are investigating agent architectures and methods, that implement reinforcement learning from scratch based on the principle of "collect and infer" and investigate their application to challenging control and robotic domains. more...

Bringing reinforcement learning (RL) methods from an appealing academic concept closer to real world control applications was one of my major research goals from 1994 onwards. One key focus was to improve data-efficiency by massive re-use of stored transition data, which lead to NFQ and a variant for continuous actions NFQCA. more...

Our research carried out in the BrainLinks BrainTools Excellence Cluster at the University of Freiburg focused around the question of how to use Machine Learning to interact with the (human) brain. more...

Brainstormers

(1998 - 2008)

RoboCup is an international research initiative intending to expedite AI and intelligent robotics. The Brainstormers, our robotic soccer team participated in RoboCup World Championships from 1998 to 2008. The scientific goal was to prove that machine learning based agents can successfully learn skills and team play from scratch. The two teams Brainstormers Tribots (real world) and Brainstormers 2D (simulation) won 5 World Championships and several European competitions and were awarded for several scientific innovations.

more Brainstormers Tribots...

more Brainstormers 2D ...

(1992 - 1994)

Rprop was invented in 1992 as part of my master thesis on supervised learning algorithms for neural networks (supervised by Prof. Dr. Heinrich Braun). It introduces a novel type of adaptive gradient descent technique, where the update step is only based on the sign of the gradient and its temporal evolution, not its size. This leads to a very fast and very robust supervised learning algorithm for batch learning. more...

Multi Agent Production Scheduling (MAPS) is a challenging domain for the investigation of distributed, cooperative multi-agent reinforcement learning algorithms. Central idea is to enhance every machine in a job-shop setting with an intelligent agent that learns an optimal scheduling strategy from its own experience. We investigated learning strategies for distributed cooperative agents that represent their knowledge by agent-individual neural Q functions. more...

Our research on neural networks for time-series prediction lead to the successful cooperation with companies in both financial markets (Dollar/ D-Mark exchange rates, bonds, …) and for disposition systems of German newspapers and journals. more...

(2010 - 2015)

Cognit was one of the first modern AI startups with a focus on the development and application of data-efficient, self-learning intelligent control systems. It was founded in 2010 by Simone and Martin Riedmiller to push modern machine learning methods into advanced industrial applications. more...

In early days of personal computers, the only way to get computers fast enough for playing games, was to program them in assembler. It was a lot of fun to manually play the role of a compiler, that translated high level ideas into a number of hexadecimal assembler codes, that finally made ZX81 and later the ZX Spectrum a device for arcade game playing. more...

[project deprecated in 2015]

CLS2 (pronounced: clsquare) provides a standardized framework for testing (not only) Reinforcement Learning controllers on a (growing) number of different plants. more...