epistasislab / hibachi Goto Github PK

Take the best individual from the hof at the end of the run and generate the class column again. Attach that class column to the data set and save this data set separately.

One example: https://github.com/DEAP/deap/blob/master/examples/ga/nsga2.py

At the end of a hibachi run, take the best individual from the hof and save it to disk.

See this thread for an example: https://groups.google.com/forum/#!topic/deap-users/43RJ1OEuBO0

When the final data set is saved by hibachi, make sure that the data set is tab-delimited.

We're currently experiencing issues with hibachi where it seems to be overfitting on the training dataset and doesn't create a model that generalizes to other datasets. Let's try changing the fitness evaluation procedure to help promote generalization ability of the generated GP model.

1) Random subsampling

For every individual, create 10 random subsamples of 25% of the dataset with replacement (this means that one sample can potentially be chosen multiple times for the same subsample). Evaluate the individual on those 10 subsamples and use the average igsum over the 10 subsamples as the igsum portion of the fitness.

(Note: Every individual should have 10 different random subsamples, so make sure the random seed isn't reset each time.)

2) 10-fold cross-validation

For every individual, randomly divide the dataset up into 10 even "folds." Evaluate the individual on each fold separately then use the average igsum over the 10 folds as the igsum portion of the fitness.

(Note 1: This procedure is similar to 1) but guarantees that there is no overlap between the subsamples.)

(Note 2: Every individual should have 10 different folds, so make sure the random seed isn't reset each time.)

3) Adding random noise

Instead of subsampling, take the entire dataset and randomly change 10% of the feature values to another value (e.g., 0 can be changed to 1 or 2; 1 can be changed to 0 or 2; etc.). This procedure simulates noise in the dataset. Evaluate the individual on the noisy dataset. For every individual, repeat this procedure 10 times and report the average igsum over the 10 noisy copies of the dataset.

(Note: Every individual should have 10 different noisy copies of the dataset, so make sure the random seed isn't reset each time.)

If the data set does not currently have column headers, add column headers. For the class label column, make sure the column header is Class.

Steps:

1. Run hibachi with one randomly-generated data set and store the best individual along with its fitness score.

2. Test the individual on another randomly-generated data set (with the same number of features and feature types).

3. Ensure that the fitness on the new data set is relatively close to the fitness on the original data set.

4. Perform steps 1-3 several times with different data sets and different individuals.