%0 Journal Article %1 Kositsky:2002 %A Kositsky, Michael %A Barto, Andrew G. %D 2002 %J Neurocomputing %K Actor-critic Human Motor Multiple Reinforcement control; learning methods; motion; movement units; %P 889 – 895 %T The emergence of movement units through learning with noisy efferent signals and delayed sensory feedback %V 44–46 %X Rapid human arm movements often have velocity pro.les consisting of several bell-shaped acceleration–deceleration phases, sometimes overlapping in time and sometimes appearing separately. We show how such sub-movement sequences can emerge naturally as an optimal control policy is approximated by a reinforcement learning system in the face of uncertainty and feedback delay. The system learns to generate sequences ofpulse-step commands, producing fast initial sub-movements followed by several slow corrective sub-movements that often begin before the initial sub-movement has completed. These results suggest how the nervous system might e3ciently control a stochastic motor plant under uncertainty and feedback delay.

@article{Kositsky:2002, abstract = {Rapid human arm movements often have velocity pro.les consisting of several bell-shaped acceleration–deceleration phases, sometimes overlapping in time and sometimes appearing separately. We show how such sub-movement sequences can emerge naturally as an optimal control policy is approximated by a reinforcement learning system in the face of uncertainty and feedback delay. The system learns to generate sequences ofpulse-step commands, producing fast initial sub-movements followed by several slow corrective sub-movements that often begin before the initial sub-movement has completed. These results suggest how the nervous system might e3ciently control a stochastic motor plant under uncertainty and feedback delay.}, added-at = {2009-06-26T15:25:19.000+0200}, author = {Kositsky, Michael and Barto, Andrew G.}, biburl = {https://www.bibsonomy.org/bibtex/25f6d63f8a2b37c624fa670988fcab252/butz}, comment = {Lernerstruktur die submovements lernt: Actor-critic RL-System lernt 1dof plant zu kontrollieren: Update in 200ms Schritten. Ein Muskel wird durch ein konstantes 200ms andauerndes Motorcommand u aktiviert. Der Muskelzustand (länge, geschwindigkeit und beschleunigung) (200msec verzögert) wird mit CMAC in pop-code aufgelöst. Actor critique Methode lernt genaue Bewegungen und verstärkt schnelle bewegungen. Als Aktion werden 10 verschiedene Motorkomanndos ausgegeben und an den Muskel weitergereicht. Lernt schnell ans ziel zu gelangen und zeigt charakteristische geschwindigkeitsprofile}, description = {diverse cognitive systems bib}, interhash = {eb98b6d02227bbda44af1ee3ef4b8bf2}, intrahash = {5f6d63f8a2b37c624fa670988fcab252}, journal = {Neurocomputing}, keywords = {Actor-critic Human Motor Multiple Reinforcement control; learning methods; motion; movement units;}, owner = {martin}, pages = {889 – 895}, timestamp = {2009-06-26T15:25:42.000+0200}, title = {The emergence of movement units through learning with noisy efferent signals and delayed sensory feedback}, volume = {44–46}, year = 2002 }