Backend-agnostic Tree Evaluation for Genetic Programming

The explicit vectorization of the mathematical operations required for fitness calculation can dramatically increase the efficiency of tree-based genetic programming for symbolic regression. In this paper, we introduce a modern software design for the seamless integration of vectorized math libraries with tree evaluation, and we benchmark each library in terms of runtime, solution quality and energy efficiency. The latter, in particular, is an aspect of increasing concern given the growing carbon footprint of AI. With this in mind, we introduce metrics for measuring the energy usage and power draw of the evolutionary algorithm. Our results show that an optimized math backend can decrease energy usage by as much as 35% (with a proportional decrease in runtime) without any negative effects in the quality of solutions.