Troubleshooting
Resolve common input, configuration and workflow errors
Not Enough Features or Instances
Training requires at least three viable input features. Check manual column selection and the missing-feature threshold. The configured folds also need enough examples of each class. A small synthetic example does not establish model quality on a real domain.
No Retained Algorithms
Algorithms never good in training are pruned. If none remain, review the performance direction, absolute/relative mode, epsilon and data units. Do not change the threshold solely to avoid the error without confirming its scientific meaning.
Missing or Reordered Test Features
Exploration uses the fitted retained feature order. Supply those columns with the same meanings and units; ordering is reconciled. Excluded training columns are not required. Missing required features are an error.
Undefined Evaluation Metrics
An algorithm without observed test outcomes cannot be scored. NaN is the appropriate undefined rate, not zero accuracy. Predictions can still be produced without outcome labels.
Unexpected Files or Missing Exports
Use explicit model save methods for CSV, PNG, web and MAT output. Output preference flags do not trigger automatic file creation. See Outputs.
Model Unavailable After a Partial Run
A partial stage run produces raw outputs and does not imply a complete Model. Consume run_iter() to completion or finish the required stages. Checkpoints and standalone fitted-model saves serve different purposes.
Different MATLAB and Python Results
Check the performance settings, retained portfolio, projection method/dimensions, TRACE method/purity, fitted tolerance, search budget, seed, folds and parameter units first. Numerical libraries, classifiers and optimizers can still differ under the same nominal configuration. See MATLAB Compatibility.