Abstract: Overfitting is one of the important problems that restrain the application of neural network. The traditional OBD (Optima Brain Damage) algorithm uses information of second derivative to calculate sensitivities and overfitting is restrained by deleting low-sensitivity weight. However, training of the network is needed repeatedly and its efficiency is low. In this paper, the Marquardt algorithm is incorporated into the OBD algorithm,so that a new method for pruning network--Dynamic Optimal Brain Damage (DOBD) is introduced. This algorithm simplifies the network and obtains good generalization through deleting weight parameters with low sensitivities dynamically while training at the same time. An approximate method through which sensitivities can be calculated during training with little computation and a rule that uses lower limit of sensitivity to delete the unnecessary weights are presented. The new algorithm is applied to the model of estimating stabilizer gasoline RVP (Reid Vapor Pressure) for FCC (Fluidized Catalytic Cracking) unit in refinery and is compared with OBD and the non-pruning method. It is shown that DOBD can avoid overfitting effectively and has a much faster convergence speed than OBD algorithm.

摘要： 将Marquardt算法和OBD算法结合 ,提出了一种神经网络的DOBD动态剪枝算法 .通过一种近似的动态计算灵敏度的方法和灵敏度删除下限技术 ,该算法能够在训练的同时删除低灵敏度的权重以简化网络 ,避免了OBD算法的反复训练 .新算法应用于稳定汽油蒸气压预估的模型中 .计算结果表明 ,DOBD算法可以有效地克服由于网络结构过于复杂造成的过拟合现象 ,并且比OBD算法具有更快的训练收敛速度.