Abstract
This master’s thesis investigates the application of NeuroEvolution of AugmentingTopologies (NEAT) concepts to the genetic programming of quantum algorithms. The
primary motivation for this approach lies in NEAT’s potential to enable meaningful
crossover through structural matching, protect innovation via speciation, and evolve
efficient solutions by starting with simple initial structures and incrementally growing
them. The study aims to compare the performance of a novel NEAT-based method with
a traditional approach from the current literature. Both methods are implemented and
evaluated across a set of benchmark problems.
The results reveal that the traditional approach outperforms the NEAT-based method
on the Quantum Fourier Transform and Full Adder problems while the NEAT-based
approach performs better on the Deutsch-Jozsa problem, this advantage likely stems
from the use of a fixed oracle rather than inherent strengths of the NEAT framework.
Additionally, an analysis of the evolved Pareto-optimal solutions indicates that the traditional approach consistently achieves better trade-offs between performance and circuit
size. This shows that the NEAT-based approach is not able to find a more efficient
solutions. The lower performance of the NEAT-based approach is likely due to its incremental mutation strategy not being suitable for the rugged fitness landscape of quantum
circuit problems. However, further research is needed to confirm this hypothesis.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Gabriel Kronberger (Supervisor) |