1995 Volume 31 Issue 3 Pages 391-398
In order to solve the inverse problems for the system with unknown characteristics, many researchers have used a method that uses an inverse model of the target system, which is acquired by learning. The direct inverse modeling, the forward and inverse modeling and the goal directed model inversion were proposed for the acquisition. However, precise inverse models for some kind of systems cannot be obtained by these method. Furthermore, a limited scale neural networks system has only limited precision. Errors still remains in the output of the inverse model using the neural networks system. The use of the inverse model is not always a good method for solving inverse problems.
Another way to solve the inverse problem is by an iterative method. We proposed a generalized inverse model with output feedback as a model of a human nervous system solving inverse kinematics problem of a human arm. The system acquires the inverse model of the linearized model of the human arm by using neural networks and uses the inverse model as an output feedback system. The system approximates the iterative improvement of Newton's method. We call the system Output Feedback Inverse Model. However, the precision of solutions provided by the system is low.
In order to make the precision high, we improve the system and propose a new configuration of the iterative system using neural networks. We call the improved system Modified Output Feedback Inverse Model. By using a random search technique for the initial value, the proposed system provides more precise solutions than the conventional methods.
The performance of the proposed method are shown by numerical simulations.
