Genetic Algorithm with random forest study

[abduskhazi]

I ran the genetic algorithms for 100 generations. Each generation has a population of ~2700. Quick statistic

• Scoring = (initial_score - shuffled_column) score.
  • After ~77 generations
    • The tendency of the selection is to select all the features.
    • Features selected of 421
    • R^2 score ~ 0.8 (Same as with all the features selected)
  • After 100 generations
    • Features selected = 425
    • oob score = 0.7939902031839906
    • Validation r2 score = 0.8110807424250865
    • Training r2 score = 0.9715554378481472
• Scoring = (initial_score - shuffled_column) * num_features_selected.
  • After ~83 generation, features selected = 379
    • oob score = 0.7941775120952056
    • Validation r2 score = 0.8129230314498437
    • Training r2 score = 0.9719294330971964
  • After 100 generation, features selected = 394
    • oob score = 0.792100294080195
    • Validation r2 score = 0.8092712262734456
    • Training r2 score = 0.971463775102713
• Scoring = (initial_score - shuffled_column) * num_features_selected ** 2. (Execution ID = 728263141)
  • The square signal turned out to be too strong. The genetic algorithm selects no features at all after the end of 100 generations.
• With Elitism and score = (initial_score - shuffled_column) * num_features_selected.
  • The number of features selected = 406
  • Validation R2 score = 0.8013467416964914

[abduskhazi]

The correct term here is cost as we are minimizing it.