




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1、Chapter 5 Gradient Estimation and Its Effects on AdaptationIn chapter 4 we assumed that an exact measurement of the gradient vector required by the adaptive process was available at each iteration. In most applications, however, an exact measurement is not available, and an estimate based on a limit
2、ed statistical sample must be used. GRADIENT COMPONENT ESTIMATION BY DERIVATVE MEASUREMENT222,2ddvdvdvxxll=Newtons method -the first and second derivative The method of steepest decent - the first derivative2minvxxl=+The derivative are estimated numerically by taking “central differences”。 222()()2(
3、)( )()dvvdvdvvvdvxxdxddxxdxxdd+-+-+-If 0These approximations become exact as approaches zero.For quadratic performance surfaces, we have()()()()22222vvvvvxdxddldldld+-+-=()( )()()()22222222vvvvvvxdxxddldlldld+-+-+-+-=00.511.522.533.543456789101112THE PERFORMANCE PENALTY()()( )12vvvgxdxdx轾=-+-臌(v-)(v
4、+)vFor the one-weight quadratic performance function,()()()22min22min122vvvgxldldxlld轾=+-+犏臌-+=2minvxxl=+We haveIn this result we see that is constant over a given performance function.2minminPgldxx=A dimensionless measure of the effect of the gradient estimate on the adaptive adjustment, called the
5、 “perturbation” P, can further be defined in term of as follows:DERIVATIVE MEASUREMENT AND PERFORMANCE PENALTIES WITH MULTIPLE WEIGHTSA two-dimensional gradient()min0min01122min00011 101 0 1,2TV RVvv v Rvr vr vr v vxxxx=+骣=+ 桫=+00011011rrRrr轾犏=犏臌Two-dimensional derivative measurement20000220()()()2d
6、vvvdvxxdxxdld+-+-=21111221()( )()2dvvvdvxxdxxdld+-+-=That is 111000,rr When the partial derivatives of this performance surface along coordinate are measured,the normalized performance penalty in term of P0, similar,10, vv 22001101minmin, rrPPddxx=The average perturbation during the entire measureme
7、nt is given by()2001101min122rrPPPdx+=+=Let us now define a general perturbation for L+1 weights as average of the perturbations of the individual gradient component measurements as follows( )220minmin 11Lnntrace RPLLlddxx=+01LnnavLll=+2min avPd lx=VARIANCE OF THE GRADIENT ESTIMATE( )( )22TT2kkEEW R
8、WP Wxxe=+-MSE is assume to based on N samples of We first define an unbiased estimate of rth moment (矩) of 2kexke11() NrrkrkNaea=( )( )rrkrEEaea=Example: let us derive the expected fourth moment, , under the assumption that the probability density of ,is normal with mean equal to zero and standard d
9、eviation equal to 4 e ( )22/22epesep-=have( )224/244432edpdeseeaeeesp- - =蝌similar,( )222/2222edpdMSEeseeaeeep- - =蝌The variance of the moment estimate 2varrrrE As the expected squared deviation from the mean, that is,( )( )()22222211var21 rrrrrrrNNrklrklEEENaaaa aaae ea=+-=-轾=-犏臌邋Hence,()222 rkrrkl
10、rrklrEklEEEkleae eeea轾=犏臌轾=犏臌轾轾=犏犏臌臌 The result is ( )()22222221varrrrrrrNNNNNaaaaaa轾=+-犏臌-=( )2422, so, varNaaaxx-=The values of depend on how k is distributed. For example, suppose that k is distributed normally with zero mean and with a variance . The mean fourth moment is 2443eas=( )44232varNNee
11、ssxx-=So in this case,22eas=The mean second moment is From this result for the normally distribtion of k ,we might anticipate that in generalthe variance of could be expressed as( )2varKNxx=We have shown that K is 2 when k is distribute normally with zero mean. When the distribution is normal but wi
12、th nonzero mean, K is also generally somewhat less than 2.Suppose that k is uniformly distributed with zero mean and with a standard deviation of , the expected moments for even values of r are ( )33212331rrrrrpddrssaeeeeess- -骣=桫=+蝌( )224var5Nax=Table 5.1Variance of the mean-squre-error estimateTab
13、le 5.1Variance of the mean-squre-error estimateOur estimate of the corresponding gradient component is ()()()12vvvxxdxdx=+-+where,V=W-W* 。We will continue to assume the error samples (values of ) are independentHave()()()()()()22221varvarvar41 2vvvvvNxxdxddxdxdd轾 犏轾轾=+-犏臌臌 犏犏臌=+-If converged to W*,
14、(v+) and (v - ) are approximately equal to min 。2min2varvnxxd骣轾 犏=犏 犏桫犏臌%Since the values of N and are the same for the estimates of all components of the gradient vector, and since the samples of k used in all estimates were assumed to be independent, the errors in all estimates are independent and
15、 have the same variance. The covariance matrix of the estimated gradient vector at the kth iteration is accordingly given by()()2min2cov: TkkkEINxd轾轾蜒-犏犏臌臌=EFFECTS ON THE WEIGHT-VECTOR SOLUTIONThe gradient estimation noisekkkN= +We examine the effect of the noisy gradient estimate on the weight vect
16、or solution, first with Newtons method and then with the steepest-descent method. For Newtons method 1111 kkkkkkWWRWRR Nmmm-+-=-=-*kkVWW=-()1112 12kkkkkkVVVR NVR Nmmmm-+-=-=-Note111 ,QQRVQV()()()( )11111212kkkkkVVQNVNmmmm-+-=-L=-Lhave|1-2| 1, k ,.“steady-state” solution()110 12nkknnVNmm- -= =-L-()()
17、()()1100011101212 12kkknknnVVNVVNmmmmm-= =-L =-L-MFor the method of steepest descent1kkkWWm+=-* ,2WWVNRVhave()()122kkkkkkVVRVNIR VNmmm+=-+=-AgainVQV2RV =()12kkkVIVNmm+=-L-VQVNQN1()()10101212kknkknnVRVNmmm- -= =-L()1012nkknnVNmm- -= = -Lk,”steady-state” solution:1RQ Q-=L()()2min2cov: NTkkkEIxd轾轾蜒-犏犏臌
18、臌=if ( )covTkkVkE V V轾轾=犏臌臌 ()()()()222121cov12covcovcov4 1kkkkVVNNmmmm-轾轾 =-+臌臌轾+L臌L轾=臌-()()()()21121 211111111112()12TTkkkkTTkkTTTkkkkVVV VNNVNNVmmmm-=-+LL轾-L+ L犏臌For Newtons methodFor the method of steepest-descent()()()()112111111221 21 2TTTkkkkTkkTTTkkkkVVIV VINNV NN Vmmmmmm-=-L-L+轾-L+-L犏臌()()2212cov2covcovcov4kkkkVIVNNmmmm-轾轾 =-L+臌臌轾+臌轾=L -L臌 Consider()()()( )1122
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 油气田勘探实例分析考核试卷
- 健康大数据与医疗旅游服务质量评价考核试卷
- 中药材种植与中药材物流体系建设探讨考核试卷
- 保健品市场品牌忠诚度与顾客忠诚度教育影响研究考核试卷
- 绿光芒考试题及答案
- 汽车职业考试题及答案
- 内科中职考试题及答案
- 医学真实考试题及答案
- 天地伟业java面试题及答案
- 中药药效评价中的生物信息学技术探索考核试卷
- 2022年江苏省射阳中等专业学校招聘考试真题及答案
- 给搅拌站送石子合同范本
- 2023年副主任医师(副高)-学校卫生与儿少卫生(副高)考试历年真题集锦带答案
- 法律基础(第4版)PPT完整全套教学课件
- 仓管应聘求职简历表格
- 五年级下册语文期末考试学霸夺冠解密卷人教部编版含答案
- 房屋加固工程监理规划
- 一级烟草专卖管理师理论考试题库(含答案)
- von frey丝K值表完整版
- SAP月结年结用户手册精
- 碳捕集、利用与封存技术课件
评论
0/150
提交评论