抄録
Estimating some mappings by neural networks, a part of mapping properties is often known beforehand. The back propagation type neural networks, however, do not utilize this kind of knowledge about the mapping. The present paper proposes to incorporate known nonlinear functions involved in the mapping into the back propagation type neural networks in order to utilize the knowledge about the mapping. As a result, the known and unknown parts of the mapping can be learned in preorganized and unorganized layers of the neural networks respectively.
Then, the preorganized neural network is applied to inverse dynamics problems of robot manipulators. Experimental results show that the learning abilities such as convergence characteristics, generalization abilities and parameter identification can be improved compared to the conventional one by incorporating the motion equation of the manipulator into the preorganized layer.
