Movement planning through reward-based learning of a trajectory of

Transcription

Movement planning through reward-based learning of a trajectory of
Institut für Grundlagen der Informationsverarbeitung
Projektarbeit: Robotik / Maschinelles Lernen /
Neuronale Netzwerke
Movement planning through reward-based
learning of a trajectory of network states in a
recurrent network of spiking neurons
Description / Scope of work
•
One considers a recurrent network of neurons
that controls the initiation of some fixed set of
muscle synergies (movement templates). Any
trajectory of network states produces then some
movement sequence.
•
The network is set up so that is produces spontaneously some trajectories of
network states (like in „motor babbling“).
•
Trajectories that generate successful movements receive rewards, that control the
learning rate for all synapses (where more recently activated synapses receive a
higher learning rate for the moment).
•
Concrete work to be done for a Master Project: Add to existing software for
learning trajectories an interface for generating movements
•
Master Thesis: Carry out experiments with that software, where concrete
movements for the Oncilla (probably in the simulator) are learnt through
reward-based learning
•
Will be carried out in joint work with Elmar Rückert and Wolfgang Maass
Side benefits:
• Can be supported by a Forschungsbeihilfe from AMARSi
(http://www.amarsi-project.eu)
• Is likely to lead to an interesting publication
• Could provide a first step for a subsequent Phd thesis.
Contact:
DI Elmar Rückert, [email protected]
Prof. Dr. Wolfgang Maass, [email protected]
22 Januar 2013