Testing Predictive Software Systems and A Case Study of Oil Drilling

Having a wide range of applications from weather forecasts to self-driving vehicles, predictive software systems are developed to predict unknowns or future events. Their uncertainties and inaccuracies pose a grand challenge in testing due to the missing of a mechanism to validate their correctness. Metamorphic testing technique is used to derive effective metamorphic relations for testing three commonly-used systems which rely on multiple linear regression, neural networks and probability distribution. In addition to successful tests with random data, a case study is conducted with real-life oil drilling data where predictive models were constructed and well tested.