SGLBO vs Other Optimizers for Quantum Neural Networks

This repo contains my Python implementation of Stochastic Gradient Line Bayesian Optimization (SGLBO) and a comparison against other common optimizers (e.g., SPSA, COBYLA, Adam) when training Quantum Neural Networks (QNNs) under different noise conditions.

Based on the method from:
📄 Stochastic Gradient Line Bayesian Optimization for Efficient Noise-Robust Optimization of Parameterized Quantum Circuits
Shiro Tamiya, Hayata Yamasaki (2021)

What this project does

• Implements SGLBO from scratch for variational quantum circuits.
• Runs experiments on QNNs (via Qiskit Machine Learning).
• Compares SGLBO against other optimizers across noiseless and noisy simulation levels.
• Includes a reusable experiment framework so datasets, optimizers, and noise profiles can be swapped in without rewriting code.
• Prepares for eventual runs on IBM Quantum hardware.

Why

Training quantum neural networks is hard because quantum hardware is noisy and measurement results are uncertain. Most classical optimizers struggle in this setting and often get stuck. SGLBO is a Bayesian optimization method that adapts more intelligently to noise. This project tests whether it performs better than other common optimizers when training QNNs.

Status

• SGLBO implemented and running.
• SPSA comparison implemented.
• Adding more optimizers (Adam, COBYLA).
• Preparing for real-hardware tests.

Notes

If anything breaks… pretend you didn’t see it (or open an issue 😔).