TDD.txt

Final goals

    1. Working ELM, easy to use
    2. All types of neurons in one model
    3. BLUE solution and GPU acceleration
    4. Model structure selection
    5. Classification support, class weighting
    6. Available from Matlab
    7. Timeseries support
    8. Distributed implementation


######################################################################

Notes

    1. Universal neurons with 2 inputs - weight and bias,
        suitable for RBF functions
    2. 


List of desired behaviour

    1. Solve XOR problem with 'sigm' ELM
    2. Solve Sine problem with 'sigm' ELM
    3. Handle inputs normalization (parameters are given)
    6. Read data from files: text and binary Numpy
    7. Save and load ELM model

######################################################################


Acceptance testing

    1. Solve XOR problem with one neuron
    2. Approximate sine function with ELM
    3. Run on all datasets from OP-ELM paper


Test SLFN correctness

    Can load whatever Numpy matrix data

    1. Send non-Numpy inputs, raise error
    2. Send non-Numpy targets, raise error
    3. Send 1-dim inputs, correct usage
    4. Send 1-dim outputs, correct usage
    5. Send inputs of different dimensionality, raise error
    6. Send targets of different dimensionality, raise error
    7. Send all-zero inputs, does not fail
    8. Send all-zero targets, does not fail
    9. Train with no neurons, raise error
    10. X and T have different number of samples, raise error
    11. Cannot have more linear neurons than input features
    12. Linear neurons initialize an identity matrix
    13. Add linear neurons, got them
    14. Add sigm neurons, got them
    15. Add tanh neurons, got them
    16. Add rbf_l1 neurons, got them
    17. Add rbf_l2 neurons, got them
    18. Add rbf_linf neurons, got them
    19. Add custom ufunc neurons, got them
    20. Add two types of neurons
    21. Add one type of neurons twice
    22. Init with bias
    23. Init with W
    24. Init without bias and W, both non-zero


Test simple ELM performance

    Run all benchmarks from the paper succesfully
    1. Classification_Iris
    2. Classification_Pima Indians Diabetes
    3. Classification_Wine 
    4. Classification_Wisconsin Breast Cancer
    5. Regression_Abalone
    6. Regression_Ailerons
    7. Regression_Auto price
    8. Regression_Bank
    9. Regression_Boston
    10. Regression_Breast cancer
    11. Regression_Computer
    12. Regression_CPU
    13. Regression_Elevators
    14. Regression_Servo
    15. Regression_Stocks


Test data loader

    3a. Reshape 1-dim X and add bias
    3b. Reshape 1-dim Y
    3c. Test encoder 1-dim
    3d. Test encoder 2-dim
    3e. Test encoder string
    3f. Test decoder 1-dim
    3g. Test decoder string
    3h. Classification - transform integer classes into one-out-of-all code
    3i. Classification - transform anything to classes
    3j. Get mean and std of training data
    3k. Read text files with different delimiters
    3l. Batch parameter works
    3m. Data loader gets number of inputs
    3n. Data loader gets number of targets
    3o. Data loader gets number of targets for classification
    3p. Data features consisting of 0, 1 and -1 are not normalized


Test model selection

    1. Error functino works for regression
    2. Error function works for classification
    3. Pruning works with one neuron type
    4. Pruning works with several neuron types

    5. Run regression with a validation set
    6. Run classification with a validation set
    7. Validation with multiple types of neurons