Improving Neural Network Reliability for Dynamic System Modeling and Control Optimization Through the Use of Confidence Measures

9362155李不存在为神经网络生成全面的、统计上合理的可靠性信息的算法。神经网络的可靠性受(1)训练数据量、(2)输入新颖性、(3)数据一致性和(4)时变系统动态的影响。置信度是衡量网络可靠性的指标，它通过以下几点来衡量网络的可靠性：何时提供足够的训练数据以进行良好的泛化；何时应信任神经网络的输出；以及何时应在慢时变动态系统中进行周期性的再训练。它们还可以帮助在闭环环境中实现神经网络控制器的自动化。置信度生成算法实际上是对所有神经网络的补充，可以帮助它们与现有控制器集成到生产环境中。UNICA将为影响可靠性的四个独立因素中的每一个开发和测试置信度算法。这项研究将以既定的理论和创新思想为基础，使用来自美国铝业公司铝还原过程的人工和真实数据。首席研究员李宇春和神经网络专家麻省理工学院林肯实验室的理查德·李普曼博士将提供现实世界应用暴露和理论背景的完美结合来进行这项研究。成功的结果有望在基于神经网络的过程控制应用中具有巨大的商业潜力。*

9362155 Lee No algorithms exist that generate comprehensive, statistically sound reliability information for neural networks. Reliability of neural nets is affected by (1) the amount of training data, (2) input novelty, (3) data consistency and (4) time-varying system dynamics. Confidence measures con gauge network reliability by indication when sufficient training data has been presented for good generalization, when a neural network's output should be trusted, and when periodic retraining should occur in slow time-varying dynamic systems. They can also help automate neural network controllers in a closed-loop environment. Confidence generation algorithms complement virtually all neural nets and can help their integration with existing controllers into production environments. Unica will develop and test confidence algorithms for each of the four independent factors affecting reliability. This research will be based on established theories and innovative ideas, using artificial and real-world data from "Alcoa's aluminum reduction process. The principal investigator, Yuchun Lee, along with neural network expert, Dr. Richard Lippmann of MIT Lincoln Laboratory, provide the perfect combination of real-world application exposure and theoretical background to conduct this research. Successful results are expected to have great commercial potential for incorporation into neural network-based process control applications. ***