| 1 |
|
| 2 |
Description of large scale optimization package, Version 3.0 |
| 3 |
############################################################## |
| 4 |
Patrick Heimbach, MIT/EAPS, 02-Mar-2000 |
| 5 |
Benny Cheng, NASA/JPL, 13-Apr-2010 (parallel version) |
| 6 |
|
| 7 |
reference: |
| 8 |
######### |
| 9 |
|
| 10 |
J.C. Gilbert & C. Lemarechal |
| 11 |
Some numerical experiments with variable-storage quasi-Newton algorithms |
| 12 |
Mathematical Programming 45 (1989), pp. 407-435 |
| 13 |
|
| 14 |
flow chart |
| 15 |
########## |
| 16 |
|
| 17 |
lsopt_top |
| 18 |
| |
| 19 |
|---- check arguments |
| 20 |
|---- CALL INSTORE |
| 21 |
| | |
| 22 |
| |---- determine whether OPWARMI available: |
| 23 |
| * if no: cold start: create OPWARMI |
| 24 |
| * if yes: warm start: read from OPWARMI |
| 25 |
| create or open OPWARMD |
| 26 |
| |
| 27 |
|---- check consistency between OPWARMI and model parameters |
| 28 |
| |
| 29 |
|---- >>> if COLD start: <<< |
| 30 |
| | first simulation with f.g. xx_0; output: first ff_0, gg_0 |
| 31 |
| | set first preconditioner value xdiff_0 to 1 |
| 32 |
| | store xx(0), gg(0), xdiff(0) to OPWARMD (first 3 entries) |
| 33 |
| | |
| 34 |
| >>> else: WARM start: <<< |
| 35 |
| read xx(i), gg(i) from OPWARMD (first 2 entries) |
| 36 |
| for first warm start after cold start, i=0 |
| 37 |
| |
| 38 |
| |
| 39 |
| |
| 40 |
|---- /// if ITMAX > 0: perform optimization (increment loop index i) |
| 41 |
| ( |
| 42 |
| )---- save current values of gg(i-1) -> gold(i-1), ff -> fold(i-1) |
| 43 |
| (---- CALL LSUPDXX |
| 44 |
| ) | |
| 45 |
| ( |---- >>> if jmax=0 <<< |
| 46 |
| ) | | first optimization after cold start: |
| 47 |
| ( | | preconditioner estimated via .01*ff_0 (first guess) |
| 48 |
| ) | | dd(i-1) = -gg(i-1)*preco |
| 49 |
| ( | | |
| 50 |
| ) | >>> if jmax > 0 <<< |
| 51 |
| ( | dd(i-1) = -gg(i-1) |
| 52 |
| ) | CALL HESSUPD |
| 53 |
| ( | | |
| 54 |
| ) | |---- dd(i-1) modified via Hessian approx. |
| 55 |
| ( | |
| 56 |
| ) |---- >>> if <dd,gg> >= 0 <<< |
| 57 |
| ( | ifail = 4 |
| 58 |
| ) | |
| 59 |
| ( |---- compute step size: tact(i-1) |
| 60 |
| ) |---- compute update: xdiff(i) = xx(i-1) + tact(i-1)*dd(i-1) |
| 61 |
| ( |
| 62 |
| )---- >>> if ifail = 4 <<< |
| 63 |
| ( goto 1000 |
| 64 |
| ) |
| 65 |
| (---- if (warm start) |
| 66 |
| (---- CALL OPTLINE / LSLINE |
| 67 |
| ) | |
| 68 |
| ( | |
| 69 |
| ) | |
| 70 |
| ( |---- /// loop over simulations |
| 71 |
| ) ( |
| 72 |
| ( if (linesearch ) |
| 73 |
| ( )---- CALL SIMUL |
| 74 |
| ) ( | |
| 75 |
| ( ) |---- input: xdiff(i) |
| 76 |
| ) ( |---- output: ff(i) |
| 77 |
| ) ( reads those values from file |
| 78 |
| ( ) |
| 79 |
| ( ( |---- perform quadratic line search |
| 80 |
| ) ) |
| 81 |
| ( else |
| 82 |
| ) )---- CALL SIMUL |
| 83 |
| ) ( | |
| 84 |
| ( ) |---- input: xdiff(i) |
| 85 |
| ) ( |---- output: ff(i),gg(i) |
| 86 |
| ) ( reads those values from file |
| 87 |
| ( ) |
| 88 |
| ) (---- 1st Wolfe test: |
| 89 |
| ( ) ff(i) <= tact*xpara1*<gg(i-1),dd(i-1)> |
| 90 |
| ) ( |
| 91 |
| ( )---- 2nd Wolfe test: |
| 92 |
| ) ( <gg(i),dd(i-1)> >= xpara2*<gg(i-1),dd(i-1)> |
| 93 |
| ( ) |
| 94 |
| ) (---- >>> if 1st and 2nd Wolfe tests ok <<< |
| 95 |
| ( ) | 320: update xx: xx(i) = xdiff(i) |
| 96 |
| ) ( | |
| 97 |
| ( ) >>> else |
| 98 |
| ) ) | ifail = 7 |
| 99 |
| ) ( | set stepsize equal half of previous |
| 100 |
| ( ) | value |
| 101 |
| ( ) | |
| 102 |
| ) ( update xdiff for new simulation |
| 103 |
| ( ) |
| 104 |
| ( |
| 105 |
| ) |
| 106 |
| (---- store new values xx(i), gg(i) to OPWARMD (first 2 entries) |
| 107 |
| )---- >>> if ifail = 7,8,9 <<< |
| 108 |
| ( goto 1000 |
| 109 |
| ) |
| 110 |
| (---- compute new pointers jmin, jmax to include latest values |
| 111 |
| ) gg(i)-gg(i-1), xx(i)-xx(i-1) to Hessian matrix estimate |
| 112 |
| (---- store gg(i)-gg(i-1), xx(i)-xx(i-1) to OPWARMD |
| 113 |
| ) (entries 2*jmax+2, 2*jmax+3) |
| 114 |
| ( |
| 115 |
| )---- CALL DGSCALE |
| 116 |
| ( | |
| 117 |
| ) |---- call dostore |
| 118 |
| ( | | |
| 119 |
| ) | |---- read preconditioner of previous iteration diag(i-1) |
| 120 |
| ( | from OPWARMD (3rd entry) |
| 121 |
| ) | |
| 122 |
| ( |---- compute new preconditioner diag(i), based upon diag(i-1), |
| 123 |
| ) | gg(i)-gg(i-1), xx(i)-xx(i-1) |
| 124 |
| ( | |
| 125 |
| ) |---- call dostore |
| 126 |
| ( | |
| 127 |
| ) |---- write new preconditioner diag(i) to OPWARMD (3rd entry) |
| 128 |
| ( |
| 129 |
|---- \\\ end of optimization iteration loop |
| 130 |
| |
| 131 |
| |
| 132 |
| |
| 133 |
|---- CALL OUTSTORE |
| 134 |
| | |
| 135 |
| |---- toggle linesearch flag |
| 136 |
| |---- store gnorm0, ff(i), current pointers jmin, jmax, iterabs, |
| 137 |
| |---- searchflags to OPWARMI |
| 138 |
| |
| 139 |
| xx(i+1) needs to be computed as input for offline optimization |
| 140 |
| | |
| 141 |
| |---- CALL LSUPDXX |
| 142 |
| | | |
| 143 |
| | |---- compute dd(i), tact(i) -> xdiff(i+1) = x(i) + tact(i)*dd(i) |
| 144 |
| | |
| 145 |
| |---- CALL WRITE_CONTROL |
| 146 |
| | | |
| 147 |
| | |---- write xdiff(i+1) to special file for offline optim. |
| 148 |
| |
| 149 |
|---- print final information |
| 150 |
| |
| 151 |
O |
| 152 |
|
| 153 |
|
| 154 |
|
| 155 |
Remarks: |
| 156 |
####### |
| 157 |
|
| 158 |
- Every call to simul refers to a read procedure which |
| 159 |
reads the result of an offline forward run and/or the adjoint run |
| 160 |
itmax = 0, for cold start |
| 161 |
itmax = 1, for warm start |
| 162 |
Also, at the end, x(i+1) needs to be computed and saved |
| 163 |
to be available for the offline model and adjoint run |
| 164 |
|
| 165 |
In order to achieve minimum difference between the online and offline code |
| 166 |
xdiff(i+1) is stored to file at the end of an (offline) iteration, |
| 167 |
but recomputed identically at the beginning of the next iteration. |
| 168 |
|
| 169 |
2. Number of simulations |
| 170 |
------------------------------------------------- |
| 171 |
|
| 172 |
- nfunc: controls the maximum number of safeguarded simulations |
| 173 |
|
| 174 |
Summary: From one iteration to the next the descent direction changes. |
| 175 |
The updated control used as input for these safeguarded simulations uses the same |
| 176 |
descent direction, but different step sizes. |
| 177 |
|
| 178 |
In detail: |
| 179 |
From one iteration to the next the descent direction dd changes using |
| 180 |
the result for the adjoint vector gg of the previous iteration. |
| 181 |
In lsline the updated control xdiff(i,1) = xx(i-1) + tact(i-1,1)*dd(i-1) serves as input for |
| 182 |
a forward and adjoint model run yielding a new gg(i,1). |
| 183 |
In general, the new solution passes the 1st and 2nd Wolfe tests |
| 184 |
so xdiff(i,1) represents the solution sought: xx(i) = xdiff(i,1). |
| 185 |
If one of the two tests fails, stepsize halving is invoked to determine |
| 186 |
a new trial step aize tact(i-1,2). |
| 187 |
If more than one function call is permitted, the new step size is used together |
| 188 |
with the "old" descent direction dd(i-1) (i.e. dd is not updated using the new gg(i)), |
| 189 |
to compute a new xdiff(i,2) = xx(i-1) + tact(i-1,2)*dd(i-1) that serves as input |
| 190 |
in a new forward and adjoint run, yielding gg(i,2). |
| 191 |
If now, both Wolfe tests are successfull, the updated solution is given by |
| 192 |
xx(i) = xdiff(i,2) = xx(i-1) + tact(i-1,2)*dd(i-1). |
| 193 |
|
| 194 |
3. Double-usage of fields dd and xdiff |
| 195 |
-------------------------------------- |
| 196 |
|
| 197 |
In order to save memory both the fields dd and xdiff have a double usage. |
| 198 |
|
| 199 |
- xdiff: in lsopt_top: used as x(i) - x(i-1) for Hessian update |
| 200 |
in lsline: intermediate result for control update x = x + tact*dd |
| 201 |
|
| 202 |
- dd : in lsopt_top, lsline: descent vector, dd = -gg & hessupd |
| 203 |
in dgscale: intermediate result to compute new preconditioner |
| 204 |
|
| 205 |
4. Notice for user of old code |
| 206 |
------------------------------ |
| 207 |
|
| 208 |
Three relevant changes needed to switch to new version: |
| 209 |
|
| 210 |
(i): OPWARMI file: two variables added: |
| 211 |
gnorm0 : norm of first (cold start) gradient |
| 212 |
iabsiter: total number of iterations with respect to cold start |
| 213 |
|
| 214 |
(ii): routine names that are referenced by main_lsopt.f |
| 215 |
lsoptv1 -> lsopt_top |
| 216 |
lsline1 -> lsline |
| 217 |
|
| 218 |
(iii): parameter list of lsopt_top |
| 219 |
logical loffline included |
| 220 |
|
| 221 |
parameter file data.optim |
| 222 |
######################## |
| 223 |
|
| 224 |
The optimization is controlled by a set of parameters |
| 225 |
provided through the standard input file data.optim, |
| 226 |
which is generated within the job script. |
| 227 |
|
| 228 |
NUPDATE : max. no. of update pairs (gg(i)-gg(i-1), xx(i)-xx(i-1)) |
| 229 |
to be stored in OPWARMD to estimate Hessian |
| 230 |
[pair of current iter. is stored in (2*jmax+2, 2*jmax+3) |
| 231 |
jmax must be > 0 to access these entries] |
| 232 |
Presently NUPDATE must be > 0 |
| 233 |
(i.e. iteration without reference to previous |
| 234 |
iterations through OPWARMD has not been tested) |
| 235 |
EPSX : relative precision on xx bellow which xx should not be improved |
| 236 |
EPSG : relative precision on gg below which optimization is considered successful |
| 237 |
IPRINT : controls verbose (>=1) or non-verbose output |
| 238 |
NUMITER : always 1 |
| 239 |
ITER_NUM : index of new restart file to be created (not necessarily = NUMITER!) |
| 240 |
NFUNC : max. no. of safeguarded simulations |
| 241 |
(must be > 0) |
| 242 |
is used if step size tact is interpolated; |
| 243 |
in this case, if NFUNC > 1, a new simulation is performed with |
| 244 |
same gradient but "improved" step size |
| 245 |
FC : first guess cost function value |
| 246 |
FMIN : not used |
| 247 |
|
| 248 |
OPWARMI, OPWARMD files |
| 249 |
###################### |
| 250 |
|
| 251 |
Two files retain values of previous iterations which are |
| 252 |
used in latest iteration to update Hessian. |
| 253 |
OPWARMI: contains index settings and scalar variables |
| 254 |
OPWARMD: contains vectors |
| 255 |
|
| 256 |
Structure of OPWARMI file: |
| 257 |
------------------------- |
| 258 |
n, fc, gnorm, m, jmin, jmax, sflag, tflag, safe_iter, stepsize |
| 259 |
|
| 260 |
n = nn : no. of control variables per processor |
| 261 |
fc = ff : cost value of last iteration |
| 262 |
m = nupdate : max. no. of updates for Hessian |
| 263 |
jmin, jmax : pointer indices for OPWARMD file (cf. below) |
| 264 |
gnorm : norm of gradient gg |
| 265 |
sflag : true if linesearch is applied to next iteration |
| 266 |
tflag : true if next iteration is a safeguarded simulation |
| 267 |
safe_iter : number of safeguarded simulations made so far |
| 268 |
stepsize : value of tact stepsize |
| 269 |
|
| 270 |
Structure of OPWARMD file: |
| 271 |
------------------------- |
| 272 |
entry |
| 273 |
1 : xx(i) : control vector of latest iteration |
| 274 |
2 : gg(i) : gradient of latest iteration |
| 275 |
3 : xdiff(i), diag: preconditioning vector; (1,...,1) for cold start |
| 276 |
--- |
| 277 |
2*jmax+2 : gold = g(i) - g(i-1) for last update (jmax) |
| 278 |
2*jmax+3 : xdiff = tact * d = xx(i) - xx(i-1) for last update (jmax) |
| 279 |
|
| 280 |
if jmax = 0: cold start; no Hessian update used to compute dd |
| 281 |
if jmax > nupdate, old positions are overwritten, starting |
| 282 |
with position pair (4,5) |
| 283 |
|
| 284 |
Example 1: jmin = 1, jmax = 3, mupd = 5 |
| 285 |
|
| 286 |
1 2 3 | 4 5 6 7 8 9 empty empty |
| 287 |
|___|___|___| | |___|___| |___|___| |___|___| |___|___| |___|___| |
| 288 |
0 | 1 2 3 |
| 289 |
|
| 290 |
Example 2: jmin = 3, jmax = 7, mupd = 5 ---> jmax = 2 |
| 291 |
|
| 292 |
1 2 3 | |
| 293 |
|___|___|___| | |___|___| |___|___| |___|___| |___|___| |___|___| |
| 294 |
| 6 7 3 4 5 |
| 295 |
|
| 296 |
|
| 297 |
|
| 298 |
Error handling |
| 299 |
############## |
| 300 |
|
| 301 |
ifail | description |
| 302 |
--------+---------------------------------------------------------- |
| 303 |
< 0 | should not appear (flag indic in simul.F not used) |
| 304 |
0 | normal mode during execution |
| 305 |
1 | an input argument is wrong |
| 306 |
2 | warm start file is corrupted |
| 307 |
3 | the initial gradient is too small |
| 308 |
4 | the search direction is not a descent one |
| 309 |
5 | maximal number of iterations reached |
| 310 |
6 | maximal number of simulations reached (handled passively) |
| 311 |
7 | the linesearch failed |
| 312 |
8 | the function could not be improved |
| 313 |
9 | optline parameters wrong |
| 314 |
10 | cold start, no optimization done |
| 315 |
11 | convergence achieved within precision |
| 316 |
|
| 317 |
|
Parent Directory
| 
Revision Log
| 
Revision Graph