f75a44cac3ae7bfc8810bad5127854a131d48a9c,reagent/ope/estimators/contextual_bandits_estimators.py,DoublyRobustEstimator,evaluate,#DoublyRobustEstimator#Any#,257

Before Change

        gt_avg = RunningAverage()
        for sample in input.samples:
            log_avg.add(sample.log_reward)
            weight = (
                sample.tgt_action_probabilities[sample.log_action]
                / sample.log_action_probabilities[sample.log_action]
            )
            weight = self._weight_clamper(weight)
            dm_action_reward, dm_reward = self._calc_dm_reward(
                input.action_space, sample
            )
            tgt_avg.add((sample.log_reward - dm_action_reward) * weight + dm_reward)
            gt_avg.add(sample.ground_truth_reward)
        return EstimatorResult(
            log_avg.average, tgt_avg.average, gt_avg.average, tgt_avg.count
        )

After Change

        log_avg = RunningAverage()
        logged_vals = []
        tgt_avg = RunningAverage()
        tgt_vals = []
        gt_avg = RunningAverage()
        for sample in input.samples:
            log_avg.add(sample.log_reward)
            logged_vals.append(sample.log_reward)
            dm_action_reward, dm_reward = self._calc_dm_reward(
                input.action_space, sample
            )
            tgt_result = 0.0
            weight = 0.0
            if sample.log_action is not None:
                weight = (
                    0.0
                    if sample.log_action_probabilities[sample.log_action]
                    < PROPENSITY_THRESHOLD
                    else sample.tgt_action_probabilities[sample.log_action]
                    / sample.log_action_probabilities[sample.log_action]
                )
                weight = self._weight_clamper(weight)
                assert dm_action_reward is not None
                assert dm_reward is not None
                tgt_result += (
                    sample.log_reward - dm_action_reward
                ) * weight + dm_reward
            else:
                tgt_result = dm_reward
            tgt_avg.add(tgt_result)
            tgt_vals.append(tgt_result)
            gt_avg.add(sample.ground_truth_reward)
        (
            tgt_score,
            tgt_score_normalized,

Instances