Research Area: Learning in Multi-Agent Systems

Multi-agent environments offer many opportunities for machine learning techniques. We are especially interested in reinforcement learning of cooperative behavior. The general scenario is given by a couple of individual agents that should learn to achieve a common goal. Example applications are in robotics, where a team of robots should learn to cooperate in order to score goals (see our robotic soccer project page). Are robots able to learn such concepts like passing, running to a free position or even carrying out double passes by only being told that if they score, it's good and if they lose the ball, it's bad? Part of this work is funded within the DFG-Schwerpunktprogramm SPP 1125.

An already real-world oriented realization of learning to cooperate is process scheduling, where each learning agent controls the sequence of jobs to be processed next at 'its' machine. Success or failure is given in terms of the fulfilment of several optimization criteria. (see e.g. Riedmiller and Riedmiller, 1999.)

On the theoretical side, we are interested in developing multi-agent learning algorithms that are guaranteed to converge under certain restrictions (i.e. limited communication, limited state or action information).

Work in these directions is carried out within the DFG-funded project on Learning Algorithms for Cooperative Multi-Agent Systems.

See also our publications page for further information.