An improved conceptual mapping method for B-spline CMAC

Aiming at the problem of huge virtual memory caused by high-dimensional input, tiny quantization space and too many samples, an improved conceptual mapping method for B-spline CMAC based on multi-dimensional memory is proposed. This method can avoid the address collision without Hash mapping from virtual memory to physical memory. Compared with the conventional conceptual mapping method, the proposed method needs less virtual memory address space for mapping only a few regular address in quantization space. It has greatly improved the learning precision and generalization capability under the condition of limited physical memory. Simulation results show that B-spline CMAC with the new conceptual mapping method has a higher learning precision, faster learning speed and better generalization capability than conventional CMAC. Meanwhile, the new conceptual mapping method is better than others in the integrated performance of memory, learning and generalization capability for B-spline CMAC with the same structure.