Evolving Estimation Models for Causal Testing

Causal reasoning is a promising and increasingly popular approach for testing complex software systems that cannot be tested using conventional approaches. This involves using a causal model and previous execution data to estimate and validate causal relationships between variables. To produce accurate causal estimates and reliable test outcomes, current approaches rely on users to specify the equational relationships between variables or sacrifice explainability for automation by using black-box estimation. In this paper, we present a hybrid between genetic programming and linear regression to automatically infer human-readable non-linear equations that can be used to evaluate causal test cases. Our results show that our technique tends to produce more accurate causal estimates and more reliable test outcomes than either technique used in isolation.