Solver snopt is a nonlinear solver based on SNOPT, a sparse SQP solver
by Philip E. Gill, Walter Murray and Michael A. Saunders. SNOPT is
available from
Stanford Business Software
2680 Bayshore Parkway, Suite 304
Mountain View, CA 94043
Phone: +1-415-962-8719
Fax: +1-415-962-1869
US Academic and Government users may obtain SNOPT from Philip Gill;
contact pgill@ucsd.edu (Philip Gill) to see about getting source for
library forms of SNOPT.
It is most convenient to invoke snopt with AMPL's solve command:
ampl: option solver snopt;
ampl: solve;
but snopt can also be run separately, with invocation syntax
snopt stub [-AMPL] [keywd=[value] ...]
in which stub comes from AMPL's write command. For example,
> ampl -obfoo foo.mod foo.dat
> snopt foo
demonstrates running snopt separately; the ampl invocation writes file
foo.nl (stub = "foo"), which snopt reads.
Command-line arguments to snopt either have the form
keywd=
or
keywd=value
where keywd is one of the key words described below. Alternatively,
you can invoke snopt the way AMPL's solve command does, i.e.,
snopt stub -AMPL [keywd=value ...]
where stub was specified in
ampl -obstub ...
or
ampl -ogstub...
Such an invocation causes snopt to read from stub.nl and to write stub.sol.
-----------------
Controlling snopt
-----------------
Snopt reads keywords and values from the environment (shell) variable
snopt_options and from the command line. Execute
snopt -?
or (if your shell requires ? to be quoted)
snopt '-?'
for a summary of snopt usage and
snopt -=
(or snopt '-=') for a summary of keywords peculiar to the AMPL/SNOPT
driver "snopt". You can also give SNOPT keyword phrases as keywords
if you change blanks within them to underscores and follow them with
an = (equals sign). See the "User's Guide for SNOPT Version 7"
(http://www.cam.ucsd.edu/~peg/papers/sndoc7.pdf) for a description of
SNOPT keyword phrases. A phrase of the form nn=filename associates
Fortran unit number nn with the file named "filename"; unit 6 is the
standard output on many systems.
------------------
Sample Invocations
------------------
If you're using AMPL, just say
option solver snopt;
solve;
If you've executed, say,
ampl -objunk junk.model junk.data
then you could say
snopt junk iterations=200 feas_tol=1e-5
to force snopt to run for at most 200 iterations and use feasibility
tolerance 1e-5. With the Bourne shell, either of the invocations
snopt_options='iterations=200 feas_tol=1e-5' snopt junk
or
snopt_options='iterations=200 feas_tol=1e-6'
export snopt_options
snopt junk
would have the same effect; within AMPL, specifying
option snopt_options 'iterations=200 feas_tol=1e-5', solver snopt;
solve;
would also have this effect.
This directory has two makefile variants: makefile.u for Unix and
Linux systems, and makefile.vc for Microsoft systems and the Visual
C++ compiler. We suggest you copy the relevant variant to "makefile",
then make any necessary changes to "makefile". The makefiles assume
that $M/libsnopt.a $M/libsnprint.a $M/libblas.a (Unix systems) or
$M/snopt.lib (Microsoft systems) are libraries built from the SNOPT
source (as of 21 October 2007). The makefiles and source assume f2c
calling conventions; see solvers/README.f77 for information on using
other Fortran compilers.
For possible convenience, directory src also contains analogous
makefile variants. See the comments in them.
-----------------------
solve_result_num values
=======================
Here is a table of solve_result_num values that "snopt" can return to
an AMPL session, along with the text that appears in the associated
solve_message. The last column gives the value of the snopt output
parameter INFO, as described in the SNOPT user's guide.
Value Message INFO
0 Optimal solution found 1
1 Feasible point found 2
101 The current point cannot be improved 3
102 Requested accuracy could not be achieved 11
202 Infeasible linear constraints 12
203 Infeasible linear equalities 13
204 Nonlinear infeasibilities minimized 14
205 Infeasibilities minimized 21
206 Cannot satisfy nonlinear constraints 22
301 Unbounded objective 31
302 Constraint violation limit reached 32
400 Too many iterations 33
401 Major iteration limit reached 41
402 The superbasics limit is too small 42
502 Incorrect gradients from funobj 43
503 Incorrect gradients from funcon 44
508 Singular basis 51
513 Error evaluating objective at 1st feas. pt. 52
514 Error evaluating initial constraint values 61
515 Error evaluating initial objective value
517 Bad "nn=filename" assignment
522 Ill-conditioned null-space basis 62
523 Cannot proceed into undefined region 63
524 Wrong basis file dimentions 81
530 Indefinite Hessian
550 Work arrays too small 82
551 Not enough character storage 83
552 Not enough integer storage 84
553 Not enough real storage 91
554 Invalid input argument 92
555 Wrong number of basic variables 141
556 Error in basis package 142
560 bug: unexpected return from SNOPT
561 Bug: "scream" called
562 Evaluation error
563 Derivative evaluation error
564 Bug: surprise setjmp() return
565 Floating-point error
Value 530 is possible with the default qpcheck = 1 when the problem
has a quadratic objective and linear constraints.
-----------------------
Questions about this stuff? Contact dmg@ampl.com (David M. Gay).