Connectionist Modeling of Basal Ganglia Motor Circuitry

摘要

Using a self-organizing, topology-preserving, sensorimotor architecture, wedeveloped two types of neural networks that were capable of learning, without supervision, to control a simulated, three-segment robot arm with variable degrees of freedom (3,4 or 6 df). One type was an endpoint or posture-controlling network, and the other was a trajectory controller. The hidden layers in these networks consisted of both 2D and 3D lattices comprising from 729 to 1728 neurons. Through process of trial and error, all networks learned to control the positioning of th end of the robot arm within a 3D workspace. The workspace was either a hemisphere or a cube centered at the origin of the stimulated limb. When tested after training that ranged from 2000 to 12000 trials, both networks achieved relatively uniform placement accuracy throughout the workspace, the level of accuracy varying directly with the number of processing elements and asymptotically with the duration of training. The number of trials required to achieve maximum accuracy was approximately 5 times the number of neurons in the hidden layers.