A growing multi-expert structure for open-ended unsupervised learning of sensory state spaces
Matthias Kubisch · 2017
We propose a new algorithm, called Growing Multi-Expert Structure (GMES) for unsupervised learning of discrete state spaces in open-ended reinforcement learning tasks. The proposed incremental algorithm is suitable for on-line state space learning on small-scale adaptive systems due to its low computational requirements. It directly processes sequential, raw sensory data, can follow its non-stationarity, and needs only a few parameters to be tuned. The development of the state space is demonstrated in a set of RL experiments using a physically simulated pendulum with varying learning goals started from a tabula-rasa-situation. Furthermore, we show that the system benefits from intrinsic motivation by feeding back the measured learning progress as a separate source of reward to improve the system's capabilities of self-exploration.