Recent advances in computation and statistics led to an increasing use of federated models for end-to-end system test and evaluation. A federated model is a collection of interconnected models where the outputs of a model act as inputs to subsequent models. However, the process of verifying and validating federated models is poorly understood, especially when testers have limited resources, knowledge-based uncertainties, and concerns over operational realism. Testers often struggle with determining how to best allocate limited test resources for model validation. We propose a network-based representation of federated models, where the network encodes the connections between the federation of models. Nodes of the graph are given by sub-models. A directed edge from node a to node b is drawn if a inputs into b. We quantify their uncertainties through edge weights using meta-modeling and variance-based sensitivity analysis. The network-based framework allows us to propagate the uncertainties through the federated model and optimize resource allocation for validation based on the uncertainties.
Suggested Citation
Capp, Jo Anna, John T Haman, and Dhruv Patel. A Practitioner’s Framework for Federated Model Validation Resource Allocation. IDA Product ID 3001838. Alexandria, VA: Institute for Defense Analyses, 2024.