基于改进深度强化学习的倒立摆控制器设计

小车倒立摆系统是一种具有非线性、强耦合、多变量、欠驱动等特性的自然不稳定系统,倒立摆系统的稳定控制是控制理论中的典型问题。针对该种控制目标,提出了一种基于改进深度学习策略梯度算法的控制方法,控制机构采用强化学习算法作为控制策略。其中,强化学习系统由策略神经网络和基线函数神经网络共同构成,同时神经网络激活函数采用了性能更优的Swish函数,并添加了基线函数以提高训练效率。将新的算法应用于小车倒立摆系统进行仿真实验,并与经典控制算法进行比较,试验结果证明了本文算法的有效性。

基于深度强化学习的码率自适应算法研究

码率自适应(Adaptive BitRate,ABR)算法是视频客户端提高用户体验质量(Quality of Experience,QoE)的一种有效途径.针对现有ABR算法存在频繁缓冲、视频卡顿、画质较低和网络吞吐量预测不准确等问题,本文提出一种基于深度强化学习的码率自适应(Deep Reinforcement Learning based ABR,DRLA)算法.DRLA用实际网络带宽数据训练神经网络,通过收集客户端缓冲区占用率和网络吞吐量向视频服务器请求最佳码率的视频.首先,DRLA用基线函数方法优化损失函数L,用熵随机探索方法防止损失函数局部收敛;其次利用约束条件限制新旧策略的散度更新幅度提高算法的鲁棒性;最后通过置信域(trust region)优化找到最优策略,使得QoE达到最优.与现有ABR算法对比的实验结果表明:DRLA减少了训练时间,能进一步提高算法的鲁棒性和用户的QoE,并在实际环境下验证了算法的有效性.

退出盯住汇率制度的动因:基于生存分析技术的实证研究

本文利用生存分析技术对盯住汇率制度的持续期进行了实证研究。本文收集了23个国家48个样本数据。考虑到不同退出模式下经济体的不同表现,将样本国家分成了两类:贬值性退出模式的国家和非贬值性退出模式的国家,然后使用非参数方法和半参数方法,对这两类样本群体分别估计了危险率、基线危险函数和影响因素的系数。实证结果表明,两种退出模式下的决定因素存在着明显的差异。盯住汇率的持续期是决定退出模式和退出可能性的重要决定因素:贬值性退出的可能性会随着时间的推移而不断的增加,而非贬值性退出更容易发生在盯住汇率制度的早期;此外,某些决定因素在两类退出中表现迥异,比如,金融深化程度的提高会增加贬值性退出的可能性,但却降低了非贬值性退出盯住汇率制度的可能性。

Application of Radial Basis Function Network in Sensor Failure Detection

Aim To detect sensor failure in control system using a single sensor signal. Methods A neural predictor was designed based on a radial basis function network(RBFN), and the neural predictor learned the sensor signal on line with a hybrid algorithm composed of n means clustering and Kalman filter and then gave the estimation of the sensor signal at the next step. If the difference between the estimation and the actural values of the sensor signal exceeded a threshold, the sensor could be declared to have a failure. The choice of the failure detection threshold depends on the noise variance and the possible prediction error of neural predictor. Results and Conclusion\ The computer simulation results show the proposed method can detect sensor failure correctly for a gyro in an automotive engine.

A New Strategy of Integrated Control and On-line Optimization on High-purity Distillation Process

For high-purity distillation processes,it is difficult to achieve a good direct product quality control using traditional proportional-integral-differential(PID)control or multivariable predictive control technique due to some difficulties,such as long response time,many un-measurable disturbances,and the reliability and precision issues of product quality soft-sensors.In this paper,based on the first principle analysis and dynamic simulation of a distillation process,a new predictive control scheme is proposed by using the split ratio of distillate flow rate to that of bottoms as an essential controlled variable.Correspondingly,a new strategy with integrated control and on-line optimization is developed,which consists of model predictive control of the split ratio,surrogate model based on radial basis function neural network for optimization,and modified differential evolution optimization algorithm. With the strategy,the process achieves its steady state quickly,so more profit can be obtained.The proposed strategy has been successfully applied to a gas separation plant for more than three years,which shows that the strategy is feasible and effective.

Application and comparison of RNN, RBFNN and MNLR approaches on prediction of flotation column performance

Evaluation of grade and recovery plays an important role in process control and plant profitability in mineral processing operations, especially flotation. The accurate measurement or estimation of these two parameters, based on the secondary variables, is a critical issue. Data-driven modeling techniques, which entail comprehensive data analysis and implementation of machine learning methods for system forecast, provide an attractive alternative. In this paper, two types of artificial neural networks(ANNs),namely radial basis function neural network(RBFNN) and layer recurrent neural network(RNN), and also a multivariate nonlinear regression(MNLR) model were employed to predict metallurgical performance of the flotation column. The training capacity and the accuracy of these three above mentioned types of models were compared. In order to acquire data for the simulation, a case study was conducted at Sarcheshmeh copper complex pilot plant. Based on the root mean squared error and correlation coefficient values, at training and testing stages, the RNN forecasted the metallurgical performance of the flotation column better than RBF and MNLR models. The RNN could predict Cu grade and recovery with correlation coefficients of 0.92 and 0.9, respectively in testing process.

抛物线类渠道断面收缩水深的计算通式

为了得到抛物线类渠道收缩水深的显函数计算通式,该文首次提出基线函数增量法的概念并采用该方法确定函数的参数。通过对抛物线型渠道断面收缩水深的基本方程进行恒等变换,取得简单快速收敛的迭代公式;根据无量纲收缩水深与已知量综合参数之间的数值分析,采用基线函数增量法确定迭代初值函数中的参数并应用迭代理论,得到了n次抛物线型渠道收缩水深的直接计算通式。误差分析及实例计算表明,在工程常用范围即无量纲收缩水深α∈(0,0.5]范围内,在工程常用抛物线指数n∈[1,5]范围,收缩水深最大相对误差小于0.52%,在抛物线指数n∈[1,100]范围内,收缩水深的最大相对误差小于0.85%,计算公式形式简单、精度高、通用性强。

High-Order Dispersion Coefficients for Alkali-metal Atoms

High-order dispersion coefficients C9, C11, C12, and C13 for the ground-state alkali-metals were calculated by combining the 1-dependent model potential of alkali-metal atoms and linear variation method based on B-spline basis functions. The results were compared.

MINIMIZING A LINEAR FRACTIONAL FUNCTION SUBJECT TO A SYSTEM OF SUP-T EQUATIONS WITH A CONTINUOUS ARCHIMEDEAN TRIANGULAR NORM

This paper shows that the problem of minimizing a linear fractional function subject to asystem of sup-T equations with a continuous Archimedean triangular norm T can be reduced to a 0-1linear fractional optimization problem in polynomial time.Consequently,parametrization techniques,e.g.,Dinkelbach's algorithm,can be applied by solving a classical set covering problem in each iteration.Similar reduction can also be performed on the sup-T equation constrained optimization problems withan objective function being monotone in each variable separately.This method could be extended aswell to the case in which the triangular norm is non-Archimedean.

POINTED REPRESENTATIONS OFINFINITE DIMENSIONAL LIE ALGEBRAS

A contravaried bilinear pairing X on every M(ρ) × M(ρθ) is determined and it is provedthat M(ρ)is irreducible if and only if K is left nondegellerate. It is also proved that every cyclicpointed module is a quotient of some Verma-like poillted module; moreover if it is irreduciblethen it is a quotieDt of the Vermarlike poiDted module by the left kernel of some bilinearpairing K. In case the mass fUnction is symmetric, there exists a bilinear form on M(ρ). It isproved that unitals pointed modules are integrable. In addition, a characterization of the massfunctions of Kac-Moody algebras is given, which is a generalization of the finite dimensionalLie algebras case.