A Proposal ofContinuous TimeRecurrent Neural Networks with Neuromodulatory BiasforAdaptation toUn-experienced Environments

URL:http://www.ito.dis.titech.ac.jp/ Abstract: Regardless ofcomplex, unknown, anddynamically-changing environments, living creatures can recognize situated environments andbehave adaptively inreal-time. However itisimpossible toprepareoptimal motion trajectories withrespect toeverypossible situations inadvance. Thekeyconceptforrealizing theenvironmental cognition andmotor adaptation isa context-based elicitation ofconstraints which arecanalizing well-suited sensorimotor coordination. For this aim,inthis study, we proposeapolymorphic neural networks modelcalled CTRNN+NM (CTRNNwithneuromod- ulatory bias). Theproposed modelisapplied totwodimensional arm-reaching movementcontrol undervarious viscous force fields. Simulation results indicate that theproposed modelinherits highrobustness even though itissituated in unexperienced environments, whichhavesimilar rotation, butdifferent size ofviscous force, because itevolved "howto adapt" instead of"howtomove."