网络群负相关训练算法的等价形式

An Equivalent Form of Negative Correlation Training Method of Ensemble Networks

在基于负相关的神经网络群学习算法中,相关项由目标输出或网络群输出定义。本文首先证明无论相关项采用何种定义形式,个体网络的目标函数都等价于网络自身性能和群性能的线性组合,线性组合系数由给定参数和网络群规模共同确定。然后给出个体网络目标函数的统一定义,并分析了目标函数内参数的含义。该目标函数表明负相关算法实际是实现多目标优化,所以网络群算法无需负相关概念,从而在概念和实现两方面简化了负相关算法。

In the negative correlation training method of ensemble networks, the correlation item is defined by means of the target output or the ensemble output. We proved that the target function of an individual network is just a linear combination of its own performance and the ensemble performance, and the linear combination coefficients are determined by the prefixed parameter and the ensemble scale. With this proof, a unified target function is presented with clearly defined parameters. The target function reveals that the negative correlation method essentially belongs to multi-object optimization problems, and the concept of negative correlation is not necessary for the ensemble learning method, which results in a great simplification.