2026-05-19T20:12:45Zhttps://keep.lib.asu.edu/oai/requestoai:keep.lib.asu.edu:node-2018562025-07-17T20:16:33Zoai_pmh:alloai_pmh:repo_items201856
https://hdl.handle.net/2286/R.2.N.201856
http://rightsstatements.org/vocab/InC/1.0/
http://creativecommons.org/licenses/by-nc-sa/4.0
2025-05
37 pages
Stealey-Euchner, Stefan
Ng, Alexander
Forrest, Stephanie
Mathis, Cole
Barrett, The Honors College
Computer Science and Engineering Program
School of Mathematical and Statistical Sciences
Despite the impressive results demonstrated by modern neural network-based artificial
intelligence, it often acts as a black box, with this low interpretability restricting its use
in sensitive governmental and cybersecurity applications. We seek to combine these strong
developments with more human-readable programs synthesized via genetic programming to
create a generic model which can create highly interpretable solutions to a wide range of
problems.
We explore the effectiveness of two techniques in evolutionary program synthesis through
a novel WebAssembly bytecode-based format: (1) applying historical markings to construct
intuitive crossover and speciation and (2) providing evolutionary scaffolding in the fitness
function. The former builds upon the NeuroEvolution of Augmenting Topologies (NEAT)
framework, which evolves artificial neural networks for reinforcement learning tasks, and
the latter is inspired by scaffolding in natural evolution; we develop the analogous idea of
partial test cases to smooth out the fitness function by providing intermediary goals (partial
solutions) for individuals to reach.
These techniques allow us to use augmentation to build up minimal programs from
scratch, eliminating extraneous genetic information. We present PEAS (Program Evolution
through Augmenting Synthesis), a linear genetic algorithm that implements these techniques,
and show that they provide significant benefits in synthesizing solutions as demonstrated on
a simple programming task. In particular, we find that speciation is necessary for successful
evolution, as the model rarely finds solutions with unspeciated crossover alone. We explore
the potential, problems, and limitations of this type of model in extension to more general
tasks.
Evolutionary Algorithm
Genetic programming
Genetic Algorithm
Artificial Intelligence
WebAssembly
Evolutionary Scaffolding
Program Synthesis
Linear Genetic Algorithm
Interpretability
Rust
Program Synthesis through Evolution of Augmenting Topologies