Optimization schemes for learning the forward and inverse kinematic equations with neural network

摘要

Learning in networks has traditionally been posed as an optimization problem. The number of optimization variables equals the number of weights in the network. This has given neural-network-training, which usually requires iterative techniques, a reputation for being very slow. In this paper various techniques of optimizing criterion function to train neural-network (the gradient method, variable-metric, conjugate-gradient) are investigated. These techniques are modified somewhat by the use of a one dimensional search to improve robustness and to accelerate convergence. In this comparative study we used these algorithms to learn the forward and the inverse coordinate transformation of two degrees freedom (DOF) robot arm. The simulations show that the variable-metric combined with a one dimensional optimization provides a variety of benefits learning speed and minimizes the iteration number. The result shows better learning of forward and inverse kinematic robot model and significant reduction of learning time was obtained.