一、首先把两个作者 JORGE NOCEDAL 给的文件看明白sdrive.f和lbfgs.f
sdrive.f - is a simple driver
lbfgs.f - contains all the supporting routines
1、lbfgs.f文件包括LBFGS算法和运行函数等
主体函数为:
subroutine lbfgs(n, m, x, f, g, diagco, diag, iprint, eps, xtol, w, iflag)
int n, m, iprint[2], iflag;
double x[n], g[n], diag[n], w[n*(2*m+1)+2*m];
double eps, xtol;
bool diagco;
使用LBFGS方法,能够有效地解决大规模变量的问题。算法开始定义一个对角矩阵Hk0,之后应用多次迭代,每次迭代使用前M个BFGS的结果更新这个矩阵,直到得到Hk矩阵,这个矩阵近似等于海塞矩阵的逆矩阵。这里的M是用户指定的,会影响到运行中所需的存储空间大小。用户同样可以提供一个不满足默认选择的初始对角矩阵。算法的详细描述可以阅读"On the limited memory BFGS method C for large scale optimization" by D. Liu and J. Nocedal。
用户需要自己计算给定的函数F的值和梯度G,为了便于用户控制这些计算可以使用逆向交流(这个还没看到,看到再写?),通过iflag实现重复调用函数运算。
每次迭代都会定义步长,定义使用CSRCH的微变换MCVSRCH。
调用语句为call lbfgs(n, m, x, f, g, diagco, diag, iprint, eps, xtol, w, iflag);
下面详细解释参数:
n:整型变量,变量个数 n>0
m:整型变量,BFGS更新时使用的修正个数 3<=m<=7
x:双精度浮点数组变量,长度为n,初始化为可能的解得向量 X=(x1, x2,..., xn),程序结束时存储最优解结果向量
f:双精度浮点变量,初始化和iflag=1时为X带入的函数值
g:双精度浮点数组变量,长度为n,初始化呵iflag=1时为带入X后的函数的梯度
diagco:逻辑变量,如果用户希望每次迭代前提供初始化的对角矩阵设定为true,否则使用默认矩阵设为false。如果设为true, 每次迭代返回iflag=2且初始对角矩阵需要通过数组diag设置。
diag:双精度浮点数组,长度为n,如果diagco=true,初始化和iflag=2时,用户需要输入初始矩阵的数值,且每个值必须非负
iprint:整型数组变量,长度为2,用户设定
iprint[0]表示输出的频率--<0不生成输出,=0只在第一次和最后一次迭代后输出,>0每次迭代输出
iprint[1]表示输出类型----=0输出迭代次数,函数评估,函数值,归一后的梯度值和步长
=1除了以上之外,还输出最初点的变量向量和梯度向量
=2除了以上之外,还输出结果变量向量
=3除了以上之外,还输出结果梯度向量
eps:正的双精度浮点变量,用户设定,定义找到解终止迭代的误差范围 ||g||<epsmax(1,||x||);
xtol:正的双精度浮点变量,用户设定,机器精度,迭代在这个精度以下时自动终止
w:双精度浮点数组变量,长度为n(2m+1)+2m,作为LBFGS的运行空间
iflag:整型变量,如果返回结果为负表述出现错误,返回0表示无错终止,返回1用户必须给函数f和梯度g赋值,返回2用户需要提供初始对角矩阵
下面的错误信息就懒得翻译了,比较容易理解反正
The following negative values of IFLAG, detecting an error,
C are possible:
C
C IFLAG=-1 The line search routine MCSRCH failed. The
C parameter INFO provides more detailed information
C (see also the documentation of MCSRCH):
C
C INFO = 0 IMPROPER INPUT PARAMETERS.
C
C INFO = 2 RELATIVE WIDTH OF THE INTERVAL OF
C UNCERTAINTY IS AT MOST XTOL.
C
C INFO = 3 MORE THAN 20 FUNCTION EVALUATIONS WERE
C REQUIRED AT THE PRESENT ITERATION.
C
C INFO = 4 THE STEP IS TOO SMALL.
C
C INFO = 5 THE STEP IS TOO LARGE.
C
C INFO = 6 ROUNDING ERRORS PREVENT FURTHER PROGRESS.
C THERE MAY NOT BE A STEP WHICH SATISFIES
C THE SUFFICIENT DECREASE AND CURVATURE
C CONDITIONS. TOLERANCES MAY BE TOO SMALL.
C
C
C IFLAG=-2 The i-th diagonal element of the diagonal inverse
C Hessian approximation, given in DIAG, is not
C positive.
C
C IFLAG=-3 Improper input parameters for LBFGS (N or M are
C not positive).
on the driver:
程序要包括一个全局数据块lb2
common /lb3/mp, lp, gtol, stpmin, stpmax;
mp:默认为6的整型变量
lp:默认为6的整型变量
gtol:双精度浮点变量默认为0.9, 可以设为较低的数=0.1,不能小于1.D-0.4
stpmin stpmax:非负双精度浮点变量,定义行搜索的步长的上下界,默认为1.D-20~1.D+20
machine dependencies:xtol, stpmin, stpmax
general information:
直接调用:daxpy, ddot, lb1, mcsrch
输入输出:无输入,输出MP的特征信息和LP的错误信息
/****************************************
w 说明:
0~n存储梯度和其他临时信息
n+1~n+m存储scalars rho?
n+m+1~n+2m存储alpha,在计算h*g的时候使用
n+2m+1~n+2m+nm存储最后m次搜索的步骤结果
n+2m+nm+1~n+2m+2nm存储最后的梯度偏差
****************************************/
后面主要是代码可以看懂就不写了,直接找的有人已经翻译了-----------摘自http://qxred.ycool.com/post.1634524.html
DOUBLE PRECISION GTOL,ONE,ZERO,GNORM,DDOT,STP1,FTOL,STPMIN,
. STPMAX,STP,YS,YY,SQ,YR,BETA,XNORM
INTEGER MP,LP,ITER,NFUN,POINT,ISPT,IYPT,MAXFEV,INFO,
. BOUND,NPT,CP,I,NFEV,INMC,IYCN,ISCN
LOGICAL FINISH
C
SAVE
DATA ONE,ZERO/1.0D+0,0.0D+0/
C
C INITIALIZE
C ----------
C
IF(IFLAG.EQ.0) GO TO 10
GO TO (172,100) IFLAG
C IFLAG一开始为0,如果IFLAG=1表示上次MCSEARCH(172)需要计算g(a),g'(a)
C 判断是否满足wolfe conditions,IFLAG=2表示需要用户提供H,在CRF中,不需要用
C 户提供H
10 ITER= 0
IF(N.LE.0.OR.M.LE.0) GO TO 196
IF(GTOL.LE.1.D-04) THEN
C GTOL不能太小,默认0.9
IF(LP.GT.0) WRITE(LP,245)
GTOL=9.D-01
ENDIF
NFUN= 1
C 迭代次数
POINT= 0
FINISH= .FALSE.
IF(DIAGCO) THEN
DO 30 I=1,N
30 IF (DIAG(I).LE.ZERO) GO TO 195
ELSE
C 自动设置H
DO 40 I=1,N
40 DIAG(I)= 1.0D0
ENDIF
C
C THE WORK VECTOR W IS DIVIDED AS FOLLOWS:
C ---------------------------------------
C THE FIRST N LOCATIONS ARE USED TO STORE THE GRADIENT AND
C OTHER TEMPORARY INFORMATION.
C LOCATIONS (N+1)...(N+M) STORE THE SCALARS RHO.
C LOCATIONS (N+M+1)...(N+2M) STORE THE NUMBERS ALPHA USED
C IN THE FORMULA THAT COMPUTES H*G.
C LOCATIONS (N+2M+1)...(N+2M+NM) STORE THE LAST M SEARCH
C STEPS.
C LOCATIONS (N+2M+NM+1)...(N+2M+2NM) STORE THE LAST M
C GRADIENT DIFFERENCES.
C
C THE SEARCH STEPS AND GRADIENT DIFFERENCES ARE STORED IN A
C CIRCULAR ORDER CONTROLLED BY THE PARAMETER POINT.
C
C W=[f'_k,RHO,ALPHA,S,Y]
C 其中保证第一次计算得到的值放在第一个,后面的依次循环存放
C 比如在RHO_1在第二次迭代中计算得到,那存放RHO_i的地址为(k-2)%m+1
C 而y_1在第二次迭代中计算得到,那么存放Y_i的地址为IYPT+(k-2)%m+1
ISPT= N+2*M
IYPT= ISPT+N*M
DO 50 I=1,N
50 W(ISPT+I)= -G(I)*DIAG(I)
C q=-H*g(0)
GNORM= DSQRT(DDOT(N,G,1,G,1))
C |g(0)|
STP1= ONE/GNORM
C STP1=1/|g(0)|初始化第一次试探步长
C PARAMETERS FOR LINE SEARCH ROUTINE
C
FTOL= 1.0D-4
MAXFEV= 20
C
IF(IPRINT(1).GE.0) CALL LB1(IPRINT,ITER,NFUN,
* GNORM,N,M,X,F,G,STP,FINISH)
C
C --------------------
C MAIN ITERATION LOOP
C --------------------
C
80 ITER= ITER+1
C k=ITER
INFO=0
BOUND=ITER-1
IF(ITER.EQ.1) GO TO 165
C 第一次试探,用取a_l=0,a_t=STP1进入MCSEARCH搜索
IF (ITER .GT. M)BOUND=M
C BOUND=min(M,k-1)表示回退的次数,比如k=2时,只需要回退1次
YS= DDOT(N,W(IYPT+NPT+1),1,W(ISPT+NPT+1),1)
C NPT=(k-1),y_{k-1}*s_{k-1}
IF(.NOT.DIAGCO) THEN
C 无需用户给出H,lbfgs自动估计H,CRF中DIAGCO=FALSE
YY= DDOT(N,W(IYPT+NPT+1),1,W(IYPT+NPT+1),1)
DO 90 I=1,N
90 DIAG(I)= YS/YY
ELSE
IFLAG=2
RETURN
ENDIF
100 CONTINUE
IF(DIAGCO) THEN
DO 110 I=1,N
110 IF (DIAG(I).LE.ZERO) GO TO 195
ENDIF
C
C COMPUTE -H*G USING THE FORMULA GIVEN IN: Nocedal, J. 1980,
C "Updating quasi-Newton matrices with limited storage",
C Mathematics of Computation, Vol.24, No.151, pp. 773-782.
C ---------------------------------------------------------
C
CP= POINT
C CP=(k-2)%m+1,这里k>=2,CP从1开始,代表了ALPHA_{k-1},RHO_{k-1}
C 循环存放的序号,比如k=2时,第一次计算RHO,RHO_1的序号为CP=1
C 而(k-1)%m代表了s_{k},y_{k}的存放序号,比如s_1的偏移为0
IF (POINT.EQ.0) CP=M
C 如果k=m+1,那么CP=m
W(N+CP)= ONE/YS
C RHO_{k-1}=1/[y_{k-1}*s_{k-1}],计算并保存RHO_{k-1}
DO 112 I=1,N
112 W(I)= -G(I)
C q=-f'_k
CP= POINT
DO 125 I= 1,BOUND
CP=CP-1
IF (CP.EQ. -1)CP=M-1
C CP=(k-1-I)%m
C CP代表了s_{k-I},y_{k-I}的存放序号
C s_{k-I}从W(ISPT+CP*N+1)存放到W(ISPT+CP*N+N)
SQ= DDOT(N,W(ISPT+CP*N+1),1,W,1)
C s_{k-I}*y_{k-I}
INMC=N+M+CP+1
IYCN=IYPT+CP*N
W(INMC)= W(N+CP+1)*SQ
C ALPHA_{k-I}=RHO_{k-I}*s_{k-I}*y_{k-I}
CALL DAXPY(N,-W(INMC),W(IYCN+1),1,W,1)
C q=q-ALPHA_{k-I}*y_{k-I}
125 CONTINUE
C CP=POINT-BOUND=max{k-m,0}
DO 130 I=1,N
130 W(I)=DIAG(I)*W(I)
C r=H0*q
DO 145 I=1,BOUND
C FOR i=k-m,...,k-1
YR= DDOT(N,W(IYPT+CP*N+1),1,W,1)
BETA= W(N+CP+1)*YR
C BETA=RHO_{k-M-1+I}*y_{k-M-1+I}*r
INMC=N+M+CP+1
BETA= W(INMC)-BETA
ISCN=ISPT+CP*N
CALL DAXPY(N,BETA,W(ISCN+1),1,W,1)
C r=r+s_{k-M-1+I}*BETA
CP=CP+1
C CP+1是ALPHA,RHO的偏移,CP是y,s的偏移
IF (CP.EQ.M)CP=0
145 CONTINUE
C
C STORE THE NEW SEARCH DIRECTION
C ------------------------------
C r=-Hk*f'_k
DO 160 I=1,N
160 W(ISPT+POINT*N+I)= W(I)
C y_k=r
C OBTAIN THE ONE-DIMENSIONAL MINIMIZER OF THE FUNCTION
C BY USING THE LINE SEARCH ROUTINE MCSRCH
C ----------------------------------------------------
165 NFEV=0
STP=ONE
IF (ITER.EQ.1) STP=STP1
DO 170 I=1,N
170 W(I)=G(I)
172 CONTINUE
CALL MCSRCH(N,X,F,G,W(ISPT+POINT*N+1),STP,FTOL,
* XTOL,MAXFEV,INFO,NFEV,DIAG)
IF (INFO .EQ. -1) THEN
C 需要g'(a),g(a)
IFLAG=1
RETURN
ENDIF
IF (INFO .NE. 1) GO TO 190
NFUN= NFUN + NFEV
C
C COMPUTE THE NEW STEP AND GRADIENT CHANGE
C -----------------------------------------
C
NPT=POINT*N
DO 175 I=1,N
W(ISPT+NPT+I)= STP*W(ISPT+NPT+I)
C s_k
175 W(IYPT+NPT+I)= G(I)-W(I)
C y_k
POINT=POINT+1
IF (POINT.EQ.M)POINT=0
C
C TERMINATION TEST
C ----------------
C
GNORM= DSQRT(DDOT(N,G,1,G,1))
XNORM= DSQRT(DDOT(N,X,1,X,1))
XNORM= DMAX1(1.0D0,XNORM)
IF (GNORM/XNORM .LE. EPS) FINISH=.TRUE.
C
IF(IPRINT(1).GE.0) CALL LB1(IPRINT,ITER,NFUN,
* GNORM,N,M,X,F,G,STP,FINISH)
IF (FINISH) THEN
IFLAG=0
RETURN
ENDIF
GO TO 80
C
C ------------------------------------------------------------
C END OF MAIN ITERATION LOOP. ERROR EXITS.
C ------------------------------------------------------------
C
190 IFLAG=-1
IF(LP.GT.0) WRITE(LP,200) INFO
RETURN
195 IFLAG=-2
IF(LP.GT.0) WRITE(LP,235) I
RETURN
196 IFLAG= -3
IF(LP.GT.0) WRITE(LP,240)
C
C FORMATS
C -------
C
200 FORMAT(/' IFLAG= -1 ',/' LINE SEARCH FAILED. SEE'
. ' DOCUMENTATION OF ROUTINE MCSRCH',/' ERROR RETURN'
. ' OF LINE SEARCH: INFO= ',I2,/
. ' POSSIBLE CAUSES: FUNCTION OR GRADIENT ARE INCORRECT',/,
. ' OR INCORRECT TOLERANCES')
235 FORMAT(/' IFLAG= -2',/' THE',I5,'-TH DIAGONAL ELEMENT OF THE',/,
. ' INVERSE HESSIAN APPROXIMATION IS NOT POSITIVE')
240 FORMAT(/' IFLAG= -3',/' IMPROPER INPUT PARAMETERS (N OR M',
. ' ARE NOT POSITIVE)')
245 FORMAT(/' GTOL IS LESS THAN OR EQUAL TO 1.D-04',
. / ' IT HAS BEEN RESET TO 9.D-01')
RETURN
END
C
C LAST LINE OF SUBROUTINE LBFGS
C
C
2、sdrive.f包含一个简单的测试问题,函数最优解为X=(1,1...1),f=0
下面一边写出源代码一边说明
program sdrive //这里ndim即为X的维数
parameter(ndim = 2000, msave = 7, nwork = ndim*(2*msave+1)+2*msave)
double x[ndim], g[ndim], diag[ndim], w[nwork];
double f, eps, xtol, gtol, t1, t2, stpmin, stpmax;
int iprint[2], iflag, icall, n, m, mp, lp, j;
bool diagco;
external lb2
common /lb3/mp, lp, gtol, stpmin, stpmax;
n = 100;
m = 5;
iprint[0] = 1;
iprint[1] = 0;
diagco = false;
eps = 1.0D-5;
xtol = 1.0D-16;
icall = 0;
iflage = 0;
for(j=1; j<=n; j+=2){ x[j] = -1.2D0; x[j+1] = 1.D0;}
20:continue;
f = 0.D0;
for(j=1; j<=n; j+=2){
t1 = 1.D0-x[j];
t2 = 1.D1*(x[j+1]-x[j]*2);
g[j+1] = 2.D1*t2;
g[j] = -2.D0*(x[j]*g[j+1]+t1);
f=f+t1*2+t2*2;
}
call lbfgs(n, m, x, f, g, diag, iprint, eps, xtol, w, iflag);
if(flag<=0) go to 50
icall = icall+1;
if(icall>2000) go to 50;
go to 20;
50:continue;
end