Design project, Erciyes University · advisor: Prof. Dr. Bahriye Akay · 2025–2026 · grade: AA
Problem
FedAvg’s learning rate and momentum strongly affect the outcome of federated learning, but tuning them by hand is hard when data is distributed and imbalanced across clients.
Approach
- Developed ABC-FedAvg, a meta-optimization method that tunes both hyperparameters with Artificial Bee Colony, using short federated simulations as a proxy task.
- Split PathMNIST across 5 clients as non-IID data with a Dirichlet distribution (α = 0.5) and trained models in PyTorch for 20 communication rounds.
Outcome
Test accuracy rose from 78.05% to 85.28% compared with standard FedAvg, approaching centralized training. I wrote the study up in paper format.