PAGODA: An HPC Tool for Evolutionary Optimization with Parallel Evaluators

The Evolution Strategies algorithm (ES) provides robust optimization capabilities across diverse problems, but is limited by the speed at which training sets can be evaluated for each solution in the evolving population. In practice, these performance limitations have restricted feasible types of training data to low-fidelity simulations, and thus have dictated the maximum model accuracy. We introduce PAGODA (Parallel Automated Gradient-free Optimization with Dynamic Algorithms), a high performance framework that combines the power of an advanced parallel ES implementation with fully parallel workers in a novel dual-parallel architecture. Interfaced evaluator codes share memory directly with the evolutionary algorithm, eliminating intermediate file I/O for maximum scalability. PAGODA allows the user to delegate compute resources such that each candidate solution can be evaluated by multiple concurrent parallel instances of a coupled code, accelerating execution by several orders of magnitude. This enables the use of fundamentally new reference data types with higher fidelity than previously was possible. We demonstrate the first application of PAGODA by parameterizing forcefields for atomistic simulations with LAMMPS.

PRESENTER

Anderson, Calvin
calvin.anderson@gdit.com
937-499-3364

HPCMP PET (GDIT)

CO-AUTHORS

Anderson, Calvin
calvin.anderson@gdit.com

Lill, James
james.lill@gdit.com

CATEGORY

Computational Chemistry & Materials

SYSTEMS USED

Narwhal

SECRET

No