Guided Evolution with Binary Discriminators for ML Program Search

TL;DR

The paper tackles the challenge of AutoML program search in large primitive spaces by introducing an online binary discriminator that guides evolutionary search. It presents PAM-RT, a predictor-based adaptive mutation strategy that uses pairwise comparisons between a child and its parent to perform hill-climbing with learned guidance, enabling faster convergence and higher final fitness. A unified DAG-based representation for diverse ML components is mapped to a graph neural network that drives the binary predictor, trained online via a replay buffer. Across NAS-Bench-101, Nguyen symbolic regression, AutoRL, and Hero optimizer spaces, PAM-RT yields substantial speedups (e.g., 3.7x–4x) and superior performance, with ablations highlighting the advantage of binary predictors and GPS encodings for robust generalization and efficiency.

Abstract

How to automatically design better machine learning programs is an open problem within AutoML. While evolution has been a popular tool to search for better ML programs, using learning itself to guide the search has been less successful and less understood on harder problems but has the promise to dramatically increase the speed and final performance of the optimization process. We propose guiding evolution with a binary discriminator, trained online to distinguish which program is better given a pair of programs. The discriminator selects better programs without having to perform a costly evaluation and thus speed up the convergence of evolution. Our method can encode a wide variety of ML components including symbolic optimizers, neural architectures, RL loss functions, and symbolic regression equations with the same directed acyclic graph representation. By combining this representation with modern GNNs and an adaptive mutation strategy, we demonstrate our method can speed up evolution across a set of diverse problems including a 3.7x speedup on the symbolic search for ML optimizers and a 4x speedup for RL loss functions.

Figures (13)

• Figure 1: We encode a variety of ML components (learning optimizers input as Python code, NN architectures, RL loss functions, and symbolic equations) into the same computation graph representation and learn a GNN-based binary predictor over pairs of individuals to predict which graph has better performance.
• Figure 2: On all symbolic regression tasks, our method PAM-RT can provide faster convergence compared to regularized evolution. PAM-RT also outperforms other mutation strategies, Max-Pairwise and PAM.
• Figure 3: PAM-RT has better sample efficiency and higher maximum performance compared to regularized evolution on harder search spaces, Hero and AutoRL.
• Figure 4: Noisy oracle experiments on NAS-Bench-101 and Nguyen-12 show the benefit of a using a predictor with evolution. Dashed curves show regularized evolution baseline.
• Figure 5: Left: Test accuracy for regression vs binary predictor on random pairs from a fixed 10k dataset collected with regularized evolution for a range of symbolic regression tasks. Binary predictors have consistently higher test accuracy. Right: Task performance on Nguyen-12 task using different GNN architecture with GPS being the most performant architecture.