Warm and hot starts

When you solve a sequence of related problems, for example the same model with slightly different parameters, you can start each solve from the end state of the previous one. SNOPT then often needs fewer iterations. SNOPT has three start modes:

The basis records which variables and constraints were at a bound at the end of a solve. Solver concepts explains it in more detail.

Warm start with snopt

Pass the basis of the previous SnoptResult and start = "Warm". Usually you also start from the previous solution:

lb, ub = -10.0, 10.0
result1 = snopt(objective, gradient!, x0; lb = lb, ub = ub)

result2 = snopt(objective, gradient!, result1.x; lb = lb, ub = ub,
                start = "Warm", basis = result1.basis)

The two problems must have the same number of variables and the same number of constraints. The functions and the bounds can be different.

snopt throws an ArgumentError in these cases:

• start = "Warm" without basis.
• basis without start = "Warm".
• A basis from a problem with a different size.

If a solve stopped during the first evaluations, before SNOPT started (see Monitor and stop a solve), its basis contains no information. A warm start from it is then the same as a cold start.

Hot start

A hot start also reuses the LU factorization of the basis matrix and the Hessian approximation that SNOPT keeps in its workspace. These exist only in the workspace of the previous solve. snopt creates a new workspace for each call, so it rejects start = "Hot".

To use a hot start:

1. Open one workspace with initialize and build a low-level problem.
2. Solve the problem.
3. Change values in the same problem object, for example the starting point x. Do not change the sizes or the sparsity pattern.
4. Solve again with start = "Hot", before you close the workspace.

The SnoptA example in the low-level interface does these steps.