Is it possible to add an initial data point?

[datavistics]

The way Im currently using mango, I will always have a first run with good defaults. Is it possible to use this information somehow? I have quite wide ranges for my hyper parameters, and I think this would help a lot.

[sandeep-iitr]

Hi, Thanks for posting these questions.

We can enable such functionality to plugin good initial results/hyperparameters to use.
It will require us to make small changes in the inner structures so that users can give such input.
We will try to do it soon.

Currently, Mango does a few initial random evaluations. A good technique when having a very wide range of hyperparameters is to increase initial random to ~5 iterations, use parallel capabilities and set 'domain_size' a little larger value like 10000.
Note if your each training/evaluation is very costly, then use even a large 'domain_size' of ~50,000 where internally Mango will search next optimal values using the past iterations over very large regions that give better results in our production settings.

regards,
Sandeep