SBIR Phase II: Improving Neural Network Reliability for Dynamic System Modeling and Control Optimization Through the use of Confidence Measures

9625725李这个小型企业创新研究(SBIR)第二阶段项目将开发一个可以与真实世界神经网络应用程序集成的原型置信度估计模块。神经网络的可靠性受(1)输入新颖性、(2)数据一致性和(3)时变系统动态的影响。置信度度量可以通过指示神经网络的输出何时应该可信以及在缓慢时变的动态系统中何时应该进行定期再训练来衡量网络的可靠性。置信度生成算法实际上是对所有神经网络的补充，可以帮助它们集成到生产环境中。第二阶段将处理所有可靠性因素的置信度算法，以及衡量和组合置信度措施以产生单一概率值的方法。算法将在电解化学过程、金属冶炼过程、聚合物配方过程控制以及供应链库存管理的人工控制优化问题的数据上进行测试。置信度估计模块潜在地可以用于几乎所有真实世界的神经网络解决方案，以提高接受和执行的准确性。预计将在过程控制、金融、零售、保险和成像领域进行商业应用。置信度评估模块将对目前销售基于神经网络的产品的200-300家快速扩张的公司有用。***

9625725 Lee This Small Business Innovation Research (SBIR) Phase II project will develop a prototype confidence-estimator module that can be integrated with real-world neural-network applications. Reliability of neural nets is affected by (1) input novelty, (2) data consistency, and (3) time-varying system dynamics. Confidence measures can gauge network reliability by indicating when a neural network's output should be trusted and when periodic retraining should occur in slow time-varying dynamic systems. Confidence-generation algorithms complement virtually all neural nets and can help their integration into production environments. Phase II will address confidence algorithms for all reliability factors and a method for scaling and combining confidence measures to generate a single probabilistic value. Algorithms will be tested on data from an electrolytic chemical process, metal smelting process, a polymer formulation process control, and an artificial control optimization problem for supply-chain inventory management. A confidence-estimator module potentially can be used in virtually every real-world neural-network solution to improve both the acceptance and performance accuracy. Commercial applications are expected in process control, finance, retail, insurance, and imaging. The confidence-estimator module will be useful to 200-300 rapidly expanding companies currently selling neural-network-based products. ***