New Paper: Quasi-Newton Numerical Optimizer with *Knowledge-Driven* Step-Length Control on Input Parameter Space (LSL-BFGS)
27. April 2021
For a better numerical optimization of molecular systems, we (AMD group, Prof. Rarey) developed a variant of the BFGS algorithm enabling *the inclusion of arbitrary domain knowledge into the step-length selection*. The new method has substantial advantages if potential functions with steep repulsion terms have to be optimized. The method is generally applicable in all scenarios were step-length control is recommended including machine learning applications. The resulting LSL-BFGS implementation is available open source on github. For details, see the recent Journal of Computational Chemistry publication.
github-Link: https://github.com/rareylab/LSLOpt
Paper-Link: http://dx.doi.org/10.1002/jcc.26522