Towards Counterfactual Explanation and Assertion Inference for CPS Debugging

A team of researchers has introduced DeCaF, a novel AI-powered framework designed to tackle the notoriously difficult problem of debugging complex Cyber-Physical Systems (CPS). These systems, which integrate software with physical processes (like autonomous vehicles or medical devices), often fail in simulations due to intricate interactions between continuous signals and discrete events. Traditional debugging tools can only localize a problem to a component, but DeCaF goes further by using counterfactual reasoning—a "what-if" analysis—to pinpoint the exact input values and timing conditions that caused the failure.

DeCaF works by taking a failing test case and employing AI to generate minimal, necessary changes to the input signals that would make the test pass. It then generalizes these specific fixes into interpretable logical assertions, giving engineers actionable insights without needing deep internal model knowledge. The framework was tested across three CPS case studies, evaluating combinations of three counterfactual generators (like Genetic Algorithms) with two causal models (like Random Forest). The results showed that a KD-Tree Nearest Neighbors generator paired with an M5 model tree achieved the highest success rate, while a Genetic Algorithm with Random Forest offered the best balance of success and causal precision, demonstrating a robust new method for automated, explainable system validation.