直接映射低维小脑模型神经网络及在机器人传感器中的应用

提出了一种能高速度、高精度学习的低维小脑模型神经网络．模型算法采用直接权地址映射技术，将训练样本的输入量化后直接作为联想存储器中C个权的首地址，建立起输入与权的关系．经样本训练后，任意输入作为相近的两个样本间的权首地址，经过输出映射算法即可得到较精确的输出．实验表明，它学习非线性函数的精度比最新改进的CMAC高十倍以上，收敛速度则快五十倍以上，且算法简单，不会发散，学习过程要求的存储器很小，实现容易．此算法已用于机器人传感器的信号细分，取得了非常好的效果．

A Direct Weight Address Mapping LCMAC Neural Network and Its Application in Robot Sensor

A novel computational approach of the low-dimensional cerebella model articulation controller(LCMAC), with the features of significantly improved processing precision and speed of convergence on learning low-dimensional nonlinear functions, is proposed in this paper. The direct weight address mapping techniques are employed in the underlying algorithm of the model, and the relationship between the inputs and weights is established by taking the scaled inputs of training samples as the head address of C weights unit in the associative memory. After training with the samples, any input is taken as the head address of C weights between the two similar samples. The subsequent accurate outputs are obtained by output mapping algorithm. Simulation experiments show that the accuracy of learning and approximating nonlinear functions is over ten times higher, and the speed of convergence is over fifty times higher than the latest improved CMAC. The algorithm, which requires a small memory storage in learning process, is very simple and also easy to be implemented. It has been applied successfully for subdividing signals of robot sensor and the result is very satisfying.