Multi-fidelity Deep Symbolic Optimization

Felipe Leno da Silva, Jiachen Yang, Mikel Landajuela, Andre R Goncalves, Alexander Ladd, Daniel faissol, Brenden K. Petersen

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Symbolic Optimization, Multi-fidelity Symbolic Optimization, Deep Symbolic Optimization

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Although Symbolic Optimization (SO) can be used to model many challenging problems, the computational cost of evaluating large numbers of candidate solutions is intractable in many real-world domains for existing SO algorithms based on reinforcement learning (RL). While lower-fidelity surrogate models or simulations can be used to speed up solution evaluation, current methods for SO are unaware of the existence of the multiple fidelities and therefore do not natively account for the mismatch between lower and higher fidelities. We propose to explicitly reason over the multiple fidelities. For that, we introduce Multi-Fidelity Markov Decision Processes (MF-MDPs) and propose a whole new family of multi-fidelity SO algorithms that account for multiple fidelities and their associated costs. We conduct experimental evaluation in two challenging SO domains, Symbolic Regression and Antibody Optimization, and show that our methods outperform fidelity-agnostic and fidelity-aware baselines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3939