Efficient motor learning by self-organizing maps and implicit linear transformations

The paper considers training algorithms for robot control with self-organizing feature maps (SOFM). The work is based on the research of Ritter et al. (1991) who used an extended SOFM concept to make a three-joint robot arm system learn its inverse kinematics by visually supervised trial movements. Familiarity with their algorithms and simulation environment is necessary to understand the paper. The disadvantages of their approach are the large computational power needed and the necessity of very many (several thousand) trial movements. The paper presents algorithms developed to reduce the computational cost and the number of trial movements to a minimum. The most efficient algorithm only needs about one trial movement per neuron to reach an optimal training result.