Active training on the CMAC nonlinear adaptive system

The CMAC neural network presents a rigid architecture for learning and generalizing simultaneously, a limitation stressed with sparse or non-dense training datasets, and hardly solved by the current training algorithms. This paper proposes a novel training algorithm that overcomes the mentioned tradeoff. The training mechanism is based on the minimization of the energy of curvature of the output, solution based on the active deformable model theory. This leads to a cell-interaction-based internal update that preserves the efficient hashed indexing and the original learning capabilities, and delivers a higher generalization degree than the apriori embedded in the CMAC architecture. The theoretical analysis is supported with comparative results on the inverse kinematics of a robotic arm.