Ability to provide initial weights

[aquasync]

Hi there,

Thanks for this optimiser - in my tests on large scale sparse problems (upwards of 5k variables) it performs very well, better than SQOPT which was previously the best for my use case. One edge that SQOPT has though, is that I can provide initial weights (or a guess), and if these are close to the optimal solution it reduces the runtime dramatically (for my particular use case it was a factor of 10!).

The current python interface doesn't provide a way to provide initial weights - is this something that is possible given the design? I see there is the ability to do an update, but in this case (portfolio optimisation over time) all the inputs have changed (including their dimensions), but the solution will not be far from the initial weights given deliberate friction in the objective.

Thanks!

Charles

[RSchwan]

Hi Charles,

In the current version of the solver this is not possible yet. We have currently a ongoing project looking into warm-starting, but this will take probably until the end of the year. It's not clear yet if it's going to work well. Interior point methods are known to be tricky to be warm-started.

I'll keep this issue thread updated based on our findings.

Cheers, Roland

[aquasync]

Ah yes I see you mention that difficulty with warm starting interior point methods in the paper - I'll keep my fingers crossed then!