Research Area: Efficient Reinforcement Learning Algorithms

Reinforcement learning systems are computer programs that can incrementally improve themselves by the training information of success or failure. A popular example is a program that learns to balance a pole mounted on a cart by only being taught when the pole falls over or the cart hits the boundary (see the video below). Further examples are robots that learn to play soccer by themselves by only being told that scoring a goal means success or losing the ball means failure (see our robotic soccer and learning robots projects). Industrial applications include self-learning feedback controllers, for example automotive engines or thermostats, where reinforcement learning controllers can gradually improve their control performance by experience, despite the complexity and nonlinearities of the controlled systems.

Our research goal here is to develop reinforcement learning controllers that are efficient and robust. Efficiency means that they should require as little training experience as possible, in order to be directly applicable to real systems. Robust means that only few parameters are required, which are not critical to choose.

Active research areas are learning on abstract states (DFG-Project SFB 531), learning on multiple time scales (DFG-project), incorporation of a priori knowledge (DFG-project Fynesse), memory-based learning (DFG-project SPP 1125), and investigation of several function approximation schemes.

See also our publications page for further information.