Shown above is a 5-step workflow of our causal repair: (1) Extract the behaviors of a learning component as an I/O table. (2) Encode the dependency of the desired property outcome on the I/O behaviors with a Halpern-Pearl model. (3) Search for a counterfactual model value assignment, revealing an actual cause and a repair. (4) Decode the found assignment as a counterfactual component behavior. (5) Replace the original learning component with a repaired component that performs this counterfactual behavior to fix the system.

Citation:

• Pengyuan Lu, Ivan Ruchkin, Matthew Cleaveland, Oleg Sokolsky, Insup Lee. Causal Repair of Learning-Enabled Cyber-Physical Systems [ArXiv]. In Proceedings of the International Conference on Assured Autonomy (ICAA), Baltimore, MD, 2023.