Genetic Programming in Federated Aggregation: A comparison of FedGP with State of the Art Methods
| Tipo de publicação: | Inproceedings |
| Citação: | |
| Series: | Lecture Notes in Artificial Intelligence |
| Ano: | 2025 |
| Mês: | December |
| Publisher: | ICAART |
| Resumo: | The efficacy of a Federated Learning system is tightly coupled to its model-aggregation strategy. Such a dependency becomes critical when client data are non-IID. To address this challenge, we undertake a comprehensive empirical assessment of FedGP, a genetic programming based aggregation framework (previously proposed for heterogeneous settings) vs state of the art FL aggregators for non-IID data. Method. FedGP is benchmarked against three canonical baselines (i.e., FedAVG, FedPROX, and FedNOVA), using two clinically relevant image collections, PathMNIST and PneumoniaMNIST. Experiments are executed under both IID and deliberately skewed non-IID partitions. Performance is scrutinized through aggregate metrics (accuracy and F1-score), learning curve trajectories, distributional box plots, and formal significance testing. Results. In IID regimes, FedGP yields a modest yet consistent edge over the comparators. Under non-IID conditions, however, FedGP delivers marked and statistically significant gains in accuracy, F1-score, and training stability, whereas the reference methods exhibit pronounced variance and occasional collapse. Conclusions. The evidence substantiates the robustness and superior generalization of FedGP in heterogeneous federated environments. Although its genetic programming search entails additional computation, the overhead remains tractable for centrally orchestrated deployments. Future research will focus on reducing this cost through transfer-learning warm starts and parallel evaluation, extending FedGP to asynchronous or fully decentralized topologies, and incorporating agent-based orchestration for personalized model specialization. |
| Palavras-chave: | federated learning, FedGP, Genetic Programming, Models Aggregation, Multi-Agents System |
| Autores | |
| Adicionado por: | [] |
| Total mark: | 0 |
|
Anexos
|
|
|
|
|
Notas
|
|
|
|
|
|
Tópicos
|
|
|
|
|
