Using graph neural networks as surrogate models in genetic programming

Abstract

Surrogate models have been used for decades to speed up evolutionary algorithms, however, most of their uses are tailored for problems with simple individual encoding, like vectors of numbers. In this paper, we evaluate the possibility to use two different types of graph neural networks to predict the quality of a solution in tree-based genetic programming without evaluating the trees.

The proposed models are evaluated in a number of benchmarks from symbolic regression and reinforcement learning and show that GNNs can be successfully used as surrogate models for problems with a complex structure.

Keywords

• genetic programming
• graph neural networks
• surrogate models