ML Control Model Verification Beyond the Learning Range

The patent performs verification of the machine learning model's actions before actual deployment. The learning range identification unit determines the boundaries of learned behavior in advance, and the verification data acquisition unit collects test data within these boundaries beforehand. This preliminary verification process ensures that the model will not produce unintended actions during actual operation, thus maintaining safety while preserving adaptability.

The patent implements a feedback mechanism where the verification result output unit provides information about the model's predicted actions outside the learning range. This feedback loop allows operators to review and validate the model's behavior before deployment, ensuring that even adaptable actions remain within acceptable safety boundaries. The system continuously monitors and adjusts based on verification results.

The patent applies local quality by focusing verification efforts specifically on the boundaries of the learning range rather than uniformly across all possible states. The learning range identification unit determines specific boundary regions, and verification data is collected targeted at these critical areas. This localized approach maintains high safety standards while avoiding unnecessary complexity in regions where the model is already well-understood.

The patent performs verification not just at the exact boundaries but also in regions slightly beyond them. This partial excessive action ensures comprehensive safety coverage by testing not only the immediate boundaries but also adjacent regions where unexpected behavior might occur. This approach enhances safety without requiring verification of every possible state space.