TuneGridSearchCV: How does it select the best trial

RLlib Ray Tune stopping condition & comparisons
1

I used this line for TuneGridSearchCV:

TuneGridSearchCV(estimator=net, param_grid=param, cv=cv2, n_jobs=3, refit= True, return_train_score=True, Scoring = 'roc_auc', use_gpu=True)

TuneGridSearchCV is saying the best trial is average_test_score = 0.545177, and parameters = {alpha=0.75, and gamma = 3.5}

But,

  • I manually checked and found that this parameter set’s computed average test score is incorrect based on the given split 0, 1, and 2 test scores.
  • The best parameter set should be alpha = 0.3 and gamma = 3

image
image1579×478 18 KB

Could you please tell me - Am I missing something? or why Is it choosing a different parameter set as the best one with an incorrect computed averaged test score?