Abstract des Autors

Three computational problems to be solved for visually guided reaching movements, trajectory formation, coordinate transformation, and calculation for motor commands, are all ill posed in redundant biological control systems. These problems are ill posed in the sense that there exist an infinite number of possible solutions. Because humans effortlessly solve these ill-posed problems, past computational studies sought principles to resolve these difficulties. Multijoint point-to-point arm movement trajectories are characterized by roughly straight hand paths and bell-shaped speed profiles. Mathematical models that can reproduce these data have been examined. At the computational understanding level of Marr, only an optimization model can provide a solution to these theoretical and empirical questions. Kinematic and dynamic optimization principles have been proposed to account for those invariant features of human movement data so far. Experimental data has been found to support the dynamic optimization theory, which requires both forward and inverse models of the motor apparatus and the external world. We show that three optimization models, the minimum jerk model, minimum torque change model, and minimum motor command change model, can resolve these three ill-posed problems. We also introduced several neural network models proposed to solve these optimization problems. Verf.-Referat