Synthesis of Incremental Learning Architecture of Competitive Associative Neural Nets and Their Application to Control

在本研究中，我们从以下几个方面对竞争神经网络（can）进行了研究，并获得了成功的结果。高效增量学习方法的综合与分析：通过本研究，我们得到了can的渐近最优性，以及综合增量学习方法。通过与传统的反向传播网络（BPN）、径向基函数网络（RBFN）、支持向量回归（SVR）等在非线性回归中具有良好性能的神经网络的对比研究，本研究提出的新学习方法对各种非线性函数的逼近表现出最佳的非线性函数逼近性能，即使数据中存在噪声。采用新学习方法的can还应用于降雨估计、语音识别、非线性混沌预测等方面，取得了很好的效果。特别是在IEICE（美国电子、信息和通信工程师学会）举办的雨量估算竞赛中，我们使用CAN计算的结果获得了二等奖。网络在模型开关控制中的应用：我们已将上述方法应用于清洗硅片的RCA溶液的温度控制。实际的RCA清洗系统由于溶液为高浓度硫酸（H2SO4）、过氧化氢（H2O2）等，存在危险性和不稳定性，因此在对照的验证实验中很难获得良好的重复性。本研究在国际上首次建立了RCA清洗系统的热模型，并在大量的数值实验中得到应用，利用can模型切换控制器对RCA清洗系统的数值模型进行估计，得到最佳控制参数值，最终得到真实的RCA系统。我们还开发了RCA系统的实时模拟器，与未开放控制算法的商用控制器进行了比较研究，并验证了该控制器的有效性。在本研究中，特别是在RCA清洗系统的建模、实际实验以及实时仿真器的开发等方面，小松电子公司的帮助都起到了非常重要的作用。少

In this research, we have investigated the competitive neural nets (CANs) from the following facets, and obtained successful results.1. Synthesis and analysis of efficient incremental learning methods:As a result of this research, we have fond out asymptotic optimality of the CANs, and synthesized incremental learning methods. From comparative study with the conventional BPN (back-propagation nets), RBFN (radial basis function nets) and SVR (support vector regression) which is famous for its very good performance in nonlinear regression, the CANs with the new learning methods developed in this research show the best performance in nonlinear function approximation for various nonlinear functions even when the data involves noise. The CANs with the new learning methods are also applied to rainfall estimation, speech recognition, nonlinear chaos prediction, etc. and we obtained very good results. Especially, in a rainfall estimation contest held by IEICE (Institution of electronics, infor … More mation and communication engineers), our result using the CAN have honored the second prize.2. Application of the nets to model switching control:We have applied the above methods to temperature control of RCA solutions for cleaning silicon wafers. The actual RCA cleaning system is dangerous and unstable because the solutions are highly concentrated sulfuric acid (H2SO4), hydrogen peroxide (H2O2), etc., so that it is hard to obtain good repeatability in verification experiments of the control. We in this research have developed the thermal model of the RCA cleaning system, which is done for the first time in the world, and utilized in a lot of numerical experiments where the model switching controller using the CANs is applied the numerical model of the RCA cleaning system for estimating the best control parameter values, and the controller finally are the real RCA system. We have also developed a real time simulator of the RCA system for comparative studies with the commercial controllers whose control algorithms are not open, and we have verified the efficiency of the present controller.In this research, especially in modeling the RCA cleaning system, in the real experiments, and in the development of the real time simulator, a lot of helps of Komatsu Electronics Inc. have been very important. Less