Validation

Model validation in nnodely follows established practices in mechanical and control engineering, combining predictive accuracy assessment with explicit consideration of model complexity.

Validation is typically performed on a dedicated test set and can be integrated into the training workflow for continuous performance monitoring using the Validator module and the analyzeModel function.

Training Feedback and Reporting

During training and validation, nnodely provides both textual and graphical feedback through configurable visualization backends.

These tools enable users to inspect:

  • Convergence behavior

  • Loss evolution

  • Quality of model fitting

This feedback facilitates early detection of overfitting, instability, or poor generalization.

Validation results can also be exported as a structured PDF report. The report includes loss trends, learned fuzzy and parametric functions, and additional diagnostic metrics generated during analyzeModel, supporting systematic post-analysis and long-term documentation.

Domain-Specific Validation Metrics

In many applications, domain-specific performance indicators are essential for reliable model assessment. In physical and control systems, validation may involve:

  • Frequency-domain analysis

  • Stability margins

  • Control-oriented performance indices

These analyses can be performed within the validation workflow using analyzeModel with appropriate configuration of prediction horizons, closed-loop connections, and batch parameters.

These metrics are not commonly used in conventional machine learning but are critical for evaluating models in engineering contexts.

By incorporating both general-purpose and domain-specific validation criteria, nnodely ensures that trained models are evaluated according to the requirements of the target application.