摘要
采用了改进的 BP算法 ,用网络输出关于学习率的线性展开法来自适应优化学习率 ,用BFGS变尺度法来调整权重 ;对 BFGS法作了局部改进以处理可能发生的 BFGS公式中分母项为零的情况 .以实际制冷系统的热力性能参数实验数据为样本 ,对比研究了标准 BP算法、含学习率优化的最速下降法和含学习率优化的 BFGS法的学习过程 .结果表明 ,文中选用的含学习率优化的BFGS法学习效率及学习精度高 。
As a common method to train multilayer perceptron (MLP) networks in the simulation of refrigeration systems, the standard BP algorithm is lack of efficiency and adaptive capacity. So, it should be substituted by an improved BP algorithm. After analysis and comparison, the method of linear expansion of MLP's outputs with respect to learning rate was selected to optimize the learning rate adaptively, and the BFGS variable metric method was selected to tune the weights. The BFGS method is improved partly to deal with a possible case in which a denominator in BFGS formula becomes zero. Taking the experimental data of thermodynamic performance of refrigeration systems as training samples, the training course were compared, which use the standard BP algorithm, the steepest descent algorithm with learning rate optimization, and the BFGS algorithm with learning rate optimization, respectively. The results show that the BFGS algorithm with the learning rate optimization has the advantages of high learning efficiency, high approximation precision and good adaptive capacity.
出处
《上海交通大学学报》
EI
CAS
CSCD
北大核心
2001年第5期783-786,共4页
Journal of Shanghai Jiaotong University
基金
国家教委回国留学人员基金 (教外司留 [1997] 832号 )
上海市青年科技启明星计划 (沪科 [99]第 2 5 2号
上海交通大学科技发展基金 (机 A15 )
关键词
制冷系统
仿真
改进BP算法
Algorithms
Artificial intelligence
Backpropagation
Computer simulation
