From Dynamic Hebbian Learning for Oscillators to Adaptive Central Pattern Generators

摘要

In this contribution we use a model of adaptive frequency oscillators to build adaptive Central Pattern Generators (CPG). We use a network of adaptive coupled Hopf oscillators to dynamically learn any periodic signal. The signal is then encoded as a stable limit cycle in the network. The interest of this approach is that the learning is not an external optimization process but is embedded in the dynamics of the network. The learning is successful even when the teaching signal is noisy, and the encoded trajectory is stable against perturbations. Furthermore, the learned trajectory can easily be modulated in frequency or amplitude in a smooth way.