Originally posted by: JaneM

I'm running a basic MIP assignment model in which business groups are assigned to locations with the objective function minimizing costs.  I'm using OPL Studio version 12.2.

When I relax the solution so that some business groups can be assigned to more locations, the objective function value increase.  This change in feasible locations is the only one I am aware of, but I am continuing to check the code to see if something else is changing. 

The objective for the first solution is 1.85775 e+008

The objective for the relaxed solution is 1.95152 e+008 with the best node in that line at 1.93163 e+008.

I've run the model with the default ops file with a time limit of 1800 seconds.  The model solves in less than 60 seconds.

I've tried changing the absolute and relative MIP gap tolerances from the default 1.0e-6 to 1.0e-8.  The solutions are the same. 

Do you have any suggestions for tuning the ops file so that relaxed model would at least find the first solution? Any other suggestions?

I know that some people request that we send the model.  This model is too large to be sent and reviewed.

Thanks.

Hi,

some suggestions:

- Move to CPLEX 12.6.2

- Use the cloud version https://dropsolve-oaas.docloud.ibmcloud.com/software/analytics/docloud

- Solve the harder solution and then warmstart in the real problem

regards

Originally posted by: JaneM

Thanks for the reply. 

Can you recommend any documents on how to write the code for a warm start.  I've looked through what I have seen and couldn't find anything to help me set up a warm start?  

Should I be more concerned that the solutions cplex finds are local optimums?  It seems especially strange that it would get caught in a local optimum when the model had found a lower cost alternative in the more restrictive situation.  Are there any branch and bound parameters, or other parameters, that could be changed so that the algorithms wouldn't throw out the solutions that would have been lower costs.  

As well, the model "solves" very quickly to this non-optimal solution - less than a minute or two, so it is not as though it has been searching for quite a while and can't do any better.  

Thanks, 

Jane

Originally posted by: JaneM

I opened the warmstart.mod in the examples.  Is there any additional explanation for this example? 

Originally posted by: JaneM

Actually, I think I figured out that what I need is to have use the vectors.attach to assign my initial values to one of the variables, so I don't think I need any additional explanation on the warmstart unless it doesn't work.

Hi,

if you want to go through a file, you may use the example provided at

https://www.ibm.com/developerworks/community/forums/html/topic?id=a9847f2f-974b-46a4-ac93-d781f9d08304&ps=25

regards

Originally posted by: JaneM

Good news.  It was my mistake.  I had one restriction that was included in the relaxed model that wasn't in the original model.  As a result of that restriction, the relaxed model couldn't use the original solution.  No longer worried that CPLEX isn't finding a global solution on this problem.

Also, I did get the warm start working.