Deep Decentralized Reinforcement Learning is featured in Neural Networks

Decentralization – as a central characteristic in biological motor control enabling adaptive behavior – is in this study transferred to Deep Reinforcement Learning: In a decentralized architecture, multiple control units for a single six-legged walker are trained in a locomotion task. This study shows that training of such decentralized modules progressed much faster (driven by local decomposed rewards) and produced more robust behavior when tested for generalization.

For the full study see the publication website which provides detailed information on different decentralized architectures and how this affected learning as well as performance after training: Schilling, M., Melnik, A., Ohl, F.W., Ritter, H., and Hammer, B. (2021). “Decentralized Control and Local Information for Robust and Adaptive Decentralized Deep Reinforcement Learning”. Neural Networks, 144, pages 699-725.

Overview of the results of the study on DDRL.