thanks for this great project, but I have one question:

When will it be possible to use OpenOpt for optimal control problems? I have seen, there is one empty side for OCP at the project homepage so hopefully it will be released  soon ;-)

Best Regards,
Qasi

Would you do something from our Appeal, it would also improve quality of our answer.

--------------------
Regards, Dmitrey.

Given that the ASA C code is available, is there any chance that you would consider adding ASA into the list of available GLP solver?

http://www.ingber.com/#ASA

I am using OpenOpt with the ralg and scipy_lbfgsb solvers for some relatively large problems with 200—2500 parameters. Every evaluation of the objective function (or derivative) takes a couple of seconds, so that the whole optimisation runs for a week or two. Part of the calculations run on a Linux cluster, which unfortunately limits the job runtime to two days.

I would like to restart an optimisation from where it was stopped. Using the best previous solution as a starting point for a new run works fine for scipy_lbfgsb, but ralg takes very many function and derivative evaluations before it gets back on track (up to a day's worth of calculations!).

Is there a way to save the internal state of the solver and use that to restart an optimisation run from where it was stopped?

Your help would be greatly appreciated.
Ingo

I am dealing with an optimization problem where lb and ub are always available. I'd like to compare smart GLP solvers (galileo, pswarm, de) against a plain grid search within the given bounds. Is grid search available somewhere in OpenOpt? Shall I implement it myself?

Kind Regards,

Emanuele

By the way, scipy.optimize docstrings point to deprecated scikits.openopt yet.

Cheers, Alex.

I wrote an NLP in OpenOpt, which was before the discretization an optimal control problem. Everything worked nice, but it is a little bit difficult to have all the indices of x in mind when writing your OCP.

So is there a possibility to write something like:

x = [y,u] ?

Maybe the question is difficult to understand, so I paste the important parts of the NLP i have written:

# Optimal Control Problem

# minimize integral_from_0_to_1 of 0.5*u(t)**2 - y(t)
# subject to y_dot(t) = u(t) and y(0) = 1 

N  = 200
TN = 100
T  = 1.0
ht = T / TN 

# objective function:
f = lambda x: ht*((0.5*x[TN:N-1]**2).sum() - x[0:TN-1].sum()) # difficult to define cost functional

# h(x) = 0 constraints
def h(x):
    r = []
    r.append(x[0] - 1)
    for i in range(0,TN-1):
        r.append(x[i+1] - x[i] - ht*x[i+TN]) # difficult to look at the problem
    return r

# now i have to get back the real state and control for plotting
y = res.xf[0:TN-1] # this is the state
u = res.xf[TN:N-1] # this is the control
t = arange(0, T-ht, ht)

So you can see, in the cost functional and in the constraints it is difficult to handle with the indices of x.

Can you help me with this?

cheers

I have been working in an Portfolio Optimization Toolbox; currently I'm using as optimization engines the MOSEK, CPLEX and GAMS tools but, I want to use it to teach, so I need to switch into a "free optimization tool" like OpenOpt...

My toolbox takes some "raw data" and writes an standard MPS Format File (or LP Format) and this File is introduced into the Optimization Model, solved and then the solution file must be read in order to show the results in a GUI...

So, in order to implement OpenOpt as the "free optimization engine" what I need to know is how to modelate a tipical MILP problem so this would help me to write a kind of "filter" lp2OpenOpt and mps2OpenOpt to be included in the toolbox.

Because I'm new in this Forum and my knowledge of Phyton is limited, I need some help in order to convert a typical LP format (I choose LP format because it is more intuitive than the MPS) into OpenOpt format to be ready-to-run whit Python...

Can some member of the Forum give me a help?

Thank you in advance!!!!!

==================================================================
LP Format:

(All the variables are binary).

\ A typical example
minimize
obj:  + 0 b1 + 0 b2 + 0 b3 + 0 b4 + 0 b5 + 0 b6 + x7
subject to
c1:  - 5e+8 b1 - 5e+8 b2 - 5e+8 b3 - 2.3e+1 b4 - 2.2e+1 b5 - 1.9e+1 b6 + x7 = 0
c2:  + b1 + b4 = 1
c3:  + b2 + b5 = 1
c4:  + b3 + b6 = 1
bounds
0 <= b1 <= 1
0 <= b2 <= 1
0 <= b3 <= 1
0 <= b4 <= 1
0 <= b5 <= 1
0 <= b6 <= 1
x7 free
general
b1 b2 b3 b4 b5 b6
end

==================================================================
Equivalent MPS Format:

* MPS File wrote by Grid2MPS from CombBids_Server_V102Beta
*
* Version:  V.1.0.0.
* Date:     June 9, 2009
* Time:     22:00 hrs (GMT-6)
* File generated at: Sun Jun 21 21:33:43 2009   
*
NAME          Generic_MPS_File
*
* original model was minimizing
*
ROWS
N  obj
E  c1
E  c2
E  c3
E  c4
COLUMNS
    MARK1     'MARKER'                 'INTORG'
    b1        c1         -500000000.00
    b1        c2                  1
    b2        c1         -500000000.00
    b2        c3                  1
    b3        c1         -500000000.00
    b3        c4                  1
    b4        c1             -23.00
    b4        c2                  1
    b5        c1             -22.00
    b5        c3                  1
    b6        c1             -19.00
    b6        c4                  1
    MARK1     'MARKER'                 'INTEND'
    x7        obj                 1
    x7        c1                  1
RHS
    rhs       c2                  1
    rhs       c3                  1
    rhs       c4                  1
BOUNDS
UP bnd       b1                  1
UP bnd       b2                  1
UP bnd       b3                  1
UP bnd       b4                  1
UP bnd       b5                  1
UP bnd       b6                  1
FR bnd       x7     
ENDATA

A new version of OpenOffice.org (free MS Office alternative) has been recently released, this one has Python 2.6 included.

From what I know, the only one free nonlinear solver connected to OOo Calc is unconstrained and premature (as it is mentioned here: "Currently a draft implementation of Quasi-Newton with BFGS update is included to optimize unconstrained non-linear models. But this algorithm is still not very robust and needs massive improvement"), while lots of users and I would be interested in general constrained local and other NonLinear solvers, as well as in Matrix Problems.

Thus, has anybody willling to connect OpenOpt to OOo Calc?

I had published some letters like this one in OOo mail lists (I'm not affiliated with OOo), but they say there will be hardly any volunteers revealed. Since those lists require registration, I guess it's better to discuss it here.

Cheers,
JKR.