Hi, I’m solving a problem through Convex.jl & COSMO.jl. I want to build a BigFloat version of the same problem to refine the solution, and i’d like to warm start it with the low-precision solution. For the dummy problem that I expose here, it might not make sense, but for some kind of hard QP or ha…

I believe doing p = minimize(Convex.norm(y - x,2) + Convex.norm(y,1); numeric_type = BigFloat) instead might be enough.

Thanks for opening the issue. I think your MWE is not ideal, because the optimal solution of the first problem is y*=[0,0,0], which is the initialization value for a new variable. So warm starting would not do anything. However, it seems to not work even if you change the initial problem to: p = …

[image] migarstka: One issue might be that the problem that COSMO gets from Convex.jl has 8 primal variables (some of them are dummys) and we are only warm-starting 3 of them (and no duals). Then a nice check would be to try to warm start those variables as well. @ericphanson 2 questions: H…

How could i access those dummy variables values, together with the dual values, so that i can upgrade them to bigfloat ? They should all be BigFloats, I think. When you do numeric_type=BigFloat, Convex uses a MOI Model with element type BigFloat, and I think that enforces bigfloats everywhere (wh…

Indeed, it enforces the type on the data. In facts, this version of the MWE is enough: using Convex, COSMO y = Variable(3); x = [1.0,2,3]; p1 = minimize(Convex.norm(y - x,2) + 0.6 * Convex.norm(y,1)); solve!(p1, () -> COSMO.Optimizer(verbose=true)) eps = BigFloat("1e-100"); p2 = minimize(Convex.no…

[image] lrnv: if the dummys are warmstrated as requested, maybe the duals are not ? Yeah, the duals are not. It’s a missing feature in Convex.jl.

EDIT: I found a workaround. The following code works: using Convex, COSMO y = Variable(3); x = [1.0,2,3]; p1 = minimize(Convex.norm(y - x,2) + 0.6 * Convex.norm(y,1)); solve!(p1, () -> COSMO.Optimizer(verbose=true)) BT = BigFloat v_x = BT.(p1.model.model.optimizer.inner.vars.x) v_μ = BT.(p1.mode…

A less “hacky” way of warmstarting would be using the set_value! function provided by the Convex.jl package . Then, you can just set the warmstart flag to true when calling solve!: set_value!(y, y0) p2 = minimize(Convex.norm(y - x,2) + 0.6 * Convex.norm(y,1); numeric_type = BigFloat); solve!(p2, () …

Arbitrary precision refinnements with Convex.jl & COSMO.jl

Specific Domains Optimization (Mathematical)

lrnv May 26, 2021, 8:49am 3

Indeed, it makes the code run. But it does not warmstart, which correspond to the behavior we already had on this other thread. I’ll post an issue on COSMO.

Topic		Replies	Views
COSMO.jl & Convex.jl: setting intial values Optimization (Mathematical)	4	780	May 25, 2021
Warm start using Dualization with JuMP and COSMO Optimization (Mathematical) jump	5	339	November 12, 2023
Initialize a SDP solver with Convex Optimization (Mathematical)	6	287	October 7, 2024
Performance of optimization problem in a loop Optimization (Mathematical)	14	1006	September 15, 2020
What's the recommended way to solve LP problems with `BigFloat` constraints and no General Usage	1	286	February 13, 2021

Arbitrary precision refinnements with Convex.jl & COSMO.jl

Related topics