VW::cb_explore_adf::regcb Namespace Reference

Classes
struct	cb_explore_adf_regcb

Functions
LEARNER::base_learner *	setup (VW::config::options_i &options, vw &all)

Function Documentation

setup()

LEARNER::base_learner * VW::cb_explore_adf::regcb::setup	(	VW::config::options_i &	options,
		vw &	all
	)

Definition at line 229 of file cb_explore_adf_regcb.cc.

References prediction_type::action_probs, VW::config::option_group_definition::add(), VW::config::options_i::add_and_parse(), LEARNER::as_multiline(), label_type::cb, CB::cb_label, ACTION_SCORE::delete_action_scores(), vw::delete_prediction, f, finish_multiline_example(), LEARNER::init_learner(), VW::config::options_i::insert(), vw::label_type, learn(), parser::lp, LEARNER::make_base(), VW::config::make_option(), vw::p, predict(), VW::config::options_i::replace(), setup_base(), vw::trace_message, and VW::config::options_i::was_supplied().

Referenced by parse_reductions().

{
  using config::make_option;
  bool cb_explore_adf_option = false;
  bool regcb = false;
  const std::string mtr = "mtr";
  std::string type_string(mtr);
  bool regcbopt = false;
  float c0 = 0.;
  bool first_only = false;
  float min_cb_cost = 0.;
  float max_cb_cost = 0.;
  config::option_group_definition new_options("Contextual Bandit Exploration with Action Dependent Features");
  new_options
      .add(make_option("cb_explore_adf", cb_explore_adf_option)
               .keep()
               .help("Online explore-exploit for a contextual bandit problem with multiline action dependent features"))
      .add(make_option("regcb", regcb).keep().help("RegCB-elim exploration"))
      .add(make_option("regcbopt", regcbopt).keep().help("RegCB optimistic exploration"))
      .add(make_option("mellowness", c0).keep().default_value(0.1f).help("RegCB mellowness parameter c_0. Default 0.1"))
      .add(make_option("cb_min_cost", min_cb_cost).keep().default_value(0.f).help("lower bound on cost"))
      .add(make_option("cb_max_cost", max_cb_cost).keep().default_value(1.f).help("upper bound on cost"))
      .add(make_option("first_only", first_only).keep().help("Only explore the first action in a tie-breaking event"))
      .add(make_option("cb_type", type_string)
               .keep()
               .help("contextual bandit method to use in {ips,dr,mtr}. Default: mtr"));
  options.add_and_parse(new_options);

  if (!cb_explore_adf_option || !(options.was_supplied("regcb") || options.was_supplied("regcbopt")))
    return nullptr;

  // Ensure serialization of cb_adf in all cases.
  if (!options.was_supplied("cb_adf"))
  {
    options.insert("cb_adf", "");
  }
  if (type_string != mtr)
  {
    all.trace_message << "warning: bad cb_type, RegCB only supports mtr; resetting to mtr." << std::endl;
    options.replace("cb_type", mtr);
  }

  all.delete_prediction = ACTION_SCORE::delete_action_scores;

  // Set explore_type
  size_t problem_multiplier = 1;

  LEARNER::multi_learner* base = as_multiline(setup_base(options, all));
  all.p->lp = CB::cb_label;
  all.label_type = label_type::cb;

  using explore_type = cb_explore_adf_base<cb_explore_adf_regcb>;
  auto data = scoped_calloc_or_throw<explore_type>(regcbopt, c0, first_only, min_cb_cost, max_cb_cost);
  LEARNER::learner<explore_type, multi_ex>& l = LEARNER::init_learner(
      data, base, explore_type::learn, explore_type::predict, problem_multiplier, prediction_type::action_probs);

  l.set_finish_example(explore_type::finish_multiline_example);
  return make_base(l);
}