PROBABILISTIC PROGRAMMING FOR ADVANCED MACHINE LEARNING (PPAML)

PPAML

System design relies significantly on numerical predictions provided by computational models and simulations, however, an objective assessment of confidence in the predictions is a major challenge. This problem of quantifying uncertainty and performing optimization under uncertainty is of significant interest to designers and engineers of complex DoD and Commercial systems.

Our Role

MetaMorph is developing tools towards creating a robust design methodology that minimizes design iterations that occur when undue trust is placed in results obtained from deterministic simulations of mixed fidelity models. We are accomplishing that by:

  1. Automating development of Bayesian Surrogate Models for System Designs
  2. Automating methods for performing Optimization under Uncertainty
  3. Automating methods for incorporating Uncertainty in Model Libraries
  4. Scaling up Probabilistic Certificate of Correctness (PCC).
  5. Integrating Probabilistic Model Checking for Automated Verification. 
  Bayesian Analysis

Bayesian Analysis

  PCC Results

PCC Results

Partners