Machine learning in creating regressions

Question

I am researching what the best techniques are in generating test cases or regression suites using machine learning for black box testing.

Has anyone researched previously or able to help?

Answer 1

You can apply Neural Network conception for this problem. You should only consider that the accuracy of any machine learning application depends on the amount of data you use to train your algorithm.

To do that you should formalize your problem. NNs are intended to solve classification problems which usually answer the question "if an object belongs to the particular class".

Say, you can define a class "A test suitable for regression for this particular change".

Now you should prepare a training set which would consist of the input and the output that is presumed correct. Here one of the main issues is how you formalize the input.

While training you will basically feed your network with a number of train data entries. Each entry might represent the following:

some formal representation of the change (you will have to break down "a change" to a number of formal properties, flags, numbers, etc) concatenated with some formal representation of the test that fits regression class for the change encoded in that particular train data entry.

Thus, your data would look like:

CHNG_PROP1, CHNG_PROP2, .., CHNG_PROPN, TEST_PROP1, TEST_PROP2, .., TEST_PROP_N, GOOD_FOR_REGR
some_val,some_val,..,some_val,,some_val,,some_val,..,some_val,Y
some_val,some_val,..,some_val,,some_val,,some_val,..,some_val,N
some_val,some_val,..,some_val,,some_val,,some_val,..,some_val,N
some_val,some_val,..,some_val,,some_val,,some_val,..,some_val,Y
...
some_val,some_val,..,some_val,,some_val,,some_val,..,some_val,Y

Then for some new change you will:

• Break your new change and all your tests to set of properties, flags, numbers, etc. like you did for your training set
• Run your NN for that new change and for each of the test you have

NN gives you the answers in terms of probability so you will get the list of probabilities (one for each test) saying how well the particular test fits your problem.

Then you just take the tests which produce the probability more than a threashold and hence you're regression set is built.