MOSAIC model for sensorimotor learning and control

Citation
M. Haruno et al., MOSAIC model for sensorimotor learning and control, NEURAL COMP, 13(10), 2001, pp. 2201-2220
Citations number
23
Categorie Soggetti
Neurosciences & Behavoir","AI Robotics and Automatic Control
Journal title
NEURAL COMPUTATION
ISSN journal
08997667 → ACNP
Volume
13
Issue
10
Year of publication
2001
Pages
2201 - 2220
Database
ISI
SICI code
0899-7667(200110)13:10<2201:MMFSLA>2.0.ZU;2-O
Abstract
Humans demonstrate a remarkable ability to generate accurate and appropriat e motor behavior under many different and often uncertain environmental con ditions. We previously proposed a new modular architecture, the modular sel ection and identification for control (MOSAIC) model, for motor learning an d control based on multiple pairs of forward (predictor) and inverse (contr oller) models. The architecture simultaneously learns the multiple inverse models necessary for control as well as how to select the set of inverse mo dels appropriate for a given environment. It combines both feedforward and feedback sensorimotor information so that the controllers can be selected b oth prior to movement and subsequently during movement. This article extend s and evaluates the MOSAIC architecture in the following respects. The lear ning in the architecture was implemented by both the original gradient-desc ent method and the expectation-maximization (EM) algorithm. Unlike gradient descent, the newly derived EM algorithm is robust to the initial starting conditions and learning parameters. Second, simulations of an object manipu lation task prove that the architecture can learn to manipulate multiple ob jects and switch between them appropriately. Moreover, after learning, the model shows generalization to novel objects whose dynamics lie within the p olyhedra of already learned dynamics. Finally, when each of the dynamics is associated with a particular object shape, the model is able to select the appropriate controller before movement execution. When presented with a no vel shape-dynamic pairing, inappropriate activation of modules is observed followed by on-line correction.