Underlying implementation for explore_adf softmax

[paulusm]

Description

I've gone over the docs a few times, but I'm unclear as to which bandit evaluation algorithm underlies the explore_adf softmax explorer, i.e. how is the quality of an action predicted?

Link to Documentation Page

https://github.com/VowpalWabbit/vowpal_wabbit/wiki/Contextual-Bandit-algorithms

[sm-28601]

The explore_adf softmax reduction applies softmax exploration over cost-based predictions from a base learner. Lower-cost actions receive higher probabilities. A lambda parameter sharpens the distribution, and a minimum epsilon probability is enforced per action.

[paulusm]

Thanks Sahil. My question is : What is the form of the base learner model?