Callbacks with xpress in JuMP

Specific Domains Optimization (Mathematical)
,
1

Hi everyone,

I’m currently working on a MILP problem using the Xpress solver in JuMP. Some instances of this problem are taking considerably longer to solve to optimality, and I would like to implement solver-specific callbacks that dynamically adjust the solver’s runtime/time limit based on the optimality gap. Examples of rules can be:

  • For a 10% gap, allow 10 minutes before terminating if no significant improvement to the objective function is made, with a 9% improvement threshold.
  • For a 5% gap, allow 5 minutes before terminating if no significant improvement is made, with a 4% improvement threshold.
  • For a 1% gap, allow 3 minutes before terminating if no significant improvement is made, with a 0.7% improvement threshold.

I have tried using callbacks but without success. I have read about the use of solver-specific callbacks in Xpress, but I am not entirely sure if this is the best way to achieve my goal. How should I correctly set up these callbacks in JuMP? Are there any examples or code available for such callbacks, or alternative approaches to achieve the same objective?

Thank you for your assistance!

2

Are there any examples or code available for such callbacks, or alternative approaches to achieve the same objective?

Not yet, and I don’t know how to achieve this without looking into the details.

I’ve opened an issue to look into this: Add callback example to README · Issue #270 · jump-dev/Xpress.jl · GitHub

3

I’ve opened a PR to update the docs: Add callback example to README by odow · Pull Request #271 · jump-dev/Xpress.jl · GitHub

I guess you want something like this:

using JuMP, Xpress
model = direct_model(Xpress.Optimizer())
start_time, log_times = time(), Any[]
function my_callback_function(cb_data)
    prob = cb_data.model
    p_obj, p_bound = Ref{Cdouble}(), Ref{Cdouble}()
    Xpress.Lib.XPRSgetdblattrib(prob, Xpress.Lib.XPRS_MIPBESTOBJVAL, p_obj)
    Xpress.Lib.XPRSgetdblattrib(prob, Xpress.Lib.XPRS_BESTBOUND, p_bound)
    rel_gap = abs((p_obj[] - p_bound[]) / p_obj[])
    push!(log_times, (time = time() - start_time, gap = rel_gap))
    if log_times[end].gap < 0.1  # or some analysis of the gap
        Xpress.Lib.XPRSinterrupt(cb_data.model, 1234)
    end
    return
end
set_attribute(model, Xpress.CallbackFunction(), my_callback_function)