Simulation-assisted machine learning.

Area of Science:

• Computational Biology
• Machine Learning
• Systems Biology

Background:

• Predictive modeling often relies on detailed simulations or black-box machine learning.
• A gap exists where system mechanics are partially known but insufficient for high-fidelity simulations.

Purpose of the Study:

• To propose and evaluate a novel approach using approximate simulations to build kernels for machine learning.
• To enhance predictive modeling in scenarios with partial system knowledge.

Main Methods:

• Developed the simulation-based kernel (SimKern) concept.
• Utilized results from multiple simulations under various uncertainty scenarios to compute sample similarity.
• Employed these similarity measures to construct kernels for kernelized machine learning methods (e.g., Support Vector Machines).

Main Results:

• Demonstrated SimKern on four synthetic complex systems, including biologically inspired models and a network flow optimization model.
• Showed that SimKern outperforms "no-prior-knowledge" methods when training samples are scarce relative to features.
• Validated the approach's applicability across disciplines seeking predictive models with available approximate simulations.

Conclusions:

• Simulation-based kernels offer a powerful method for predictive modeling when complete system simulations are infeasible.
• SimKern effectively integrates domain knowledge from approximate simulations into machine learning frameworks.
• The SimKern approach is broadly applicable and beneficial for complex systems analysis.