1 of 1

Automating Optimization of Generator Tuning Parameters�With the HEP on HPC Partnership

Scientific Achievement

RAPIDS tools enable scaling up the synthetic generation of proton-proton collisions using the Monte Carlo code Pythia8 with DIY and automating the tuning of optimal parameters using the multifidelity optimizer Maestro with Decaf.

Significance and Impact

Modeling collisions that generate subatomic particles requires finding accurate tuning parameters for constrained and unconstrained optimization problems, where only stochastic estimates of the objective function are available from a Monte Carlo simulation.

Top: Accelerated convergence of our approach (blue color) compared with computing all Monte Carlo samples in advance for entire parameter space (orange color) shows the benefit of our method. Right: Workflow diagram executed with the Decaf workflow management system converges to the above results automatically.

Technical Approach

Simulate millions of Monte Carlo proton-proton collisions at scale using DIY
Fit a smooth response surface to a subspace (trust region) of the parameters
Compute the minimum statistical difference with experimentally observed collisions over the trust region
Refine the size, shape, and location of the trust region
Advance to the new trust region and generate more Monte Carlo collisions there
Automate entire workflow using Decaf

Maestro: Multi-fidelity Adaptive Ensemble Stochastic Trust-Region Optimization. M. Krishnamoorthy, S. Leyffer, S. Mrenna, T. Peterka, O. Yildiz. In preparation, 2023.

We automated and scaled up an optimization method for the automatic tuning of physics event generators.
Previously, an expert assigned weights to observables to represent the importance of the observable in the tune.
The human in the loop makes this approach to optimization inefficient and can potentially lead to bias.
We formulated the tuning task as a multifidelity optimization problem with a dynamic trust region of the parameter space where the search is focused.
We scaled up the generation of Monte Carlo events using DIY.
We coupled the dynamic generation of events with the optimization method using Decaf, generating batches of Monte Carlo events on the fly.

[1] Maestro: Multi-fidelity Adaptive Ensemble Stochastic Trust-Region Optimization. M. Krishnamoorthy, S. Leyffer, S. Mrenna, T. Peterka, O. Yildiz. In preparation, 2023.

[2] Yildiz, O., Dreher, M., Peterka, T.: Decaf: Decoupled Dataflows for In Situ Workflows. Book chapter, In Situ Visualization for Computational Science. Childs, H., Garth, C., Bennett, J., editors, 2022.

[3] Morozov, D., Peterka, T., Guo, H., Raj, M., Xu, J., Shen, H.-W.: IExchange: Asynchronous Communication and Termination Detection for Iterative Algorithms. Proceedings of IEEE LDAV'21 Symposium on Large Data Analysis and Visualization, New Orleans, LA, 2021.