Pareto Front Training For Multi-Objective Symbolic Optimization

Jonathan G. Faris, Conor F. Hayes, Andre R Goncalves, Kayla G. Sprenger, Daniel faissol, Brenden K. Petersen, Mikel Landajuela, Felipe Leno da Silva

Keywords: multi-objective optimization, symbolic optimization, antibody engineering

TL;DR: This manuscript presents Pareto Front Training (PFT), a multi-objective symbolic optimization algorithm that searches for optimal sequences by training a Recurrent Neural Network on the Pareto front of explored solutions thus far.

Abstract:

Although Symbolic Optimization (SO) solutions have successfully been used in applications ranging from Neural Architecture Search to Antibody Therapeutics Optimization, current SO algorithms are typically limited to using a single quality measure to search for optimal solutions. However, for many applications, solutions are more naturally described by multiple measures, e.g., a solar panel must be designed to maximize power generation while minimizing heat generation.

Herein, we propose Pareto Front Training (PFT), a SO algorithm that searches for token sequences by training a Recurrent Neural Network on the Pareto front of explored solutions. We evaluate PFT in an antibody optimization scenario using a real SARS-CoV-2 viral strain and show that PFT outperforms the baselines in terms of antibody binding quality, stability, and humanness. We hope PFT will inspire a new family of multi-objective SO algorithms and will help SO achieve varied new applications.

Type Of Paper: Full paper (max page 8)

Anonymous Submission: Anonymized submission.

Submission Number: 25