Towards a Meaningful Communication and Model Aggregation in

Federated Learning via Genetic Programming

Elia Pacioni

1,2 a

, Francisco Fern

andez De Vega

2 b

and Davide Calvaresi

1 c

University of Applied Sciences and Arts of Western Switzerland (HES-SO Valais/Wallis),

Rue de l’Industrie 23, Sion, 1950, Switzerland

Universidad de Extremadura, Av. Santa Teresa de Jornet, 38, M

erida, 06800, Spain

{elia.pacioni, davide.calvaresi}@hevs.ch, fcofdez@unex.es

Keywords:

Federated Learning, Multi-Agents System, Models Aggregation, Communication Efﬁciency, Genetic

Programming.

Abstract:

Federated Learning (FL) enables collaborative training of machine learning models while preserving client

data privacy. However, its conventional client-server paradigm presents two key challenges: (i) communica-

tion efﬁciency and (ii) model aggregation optimization. Inefﬁcient communication, often caused by transmit-

ting low-impact updates, results in unnecessary overhead, particularly in bandwidth-constrained environments

such as wireless or mobile networks or in scenarios with numerous clients. Furthermore, traditional aggre-

gation strategies lack the adaptability required for stable convergence and optimal performance. This paper

emphasizes the distributed nature of FL clients (agents) and advocates for local, autonomous, and intelligent

strategies to evaluate the signiﬁcance of their updates—such as using a “distance” metric relative to the global

model. This approach improves communication efﬁciency by prioritizing impactful updates. Additionally,

the paper proposes an adaptive aggregation method leveraging genetic programming and transfer learning to

dynamically evolve aggregation equations, optimizing the convergence process. By integrating insights from

multi-agent systems, the proposed approach aims to foster more efﬁcient and robust frameworks for decen-

tralized learning.

1 INTRODUCTION

Federated Learning (FL) is a collaborative ma-

chine learning paradigm introduced by Google in

2016 (McMahan et al., 2017). FL is designed to

train models on decentralized data while preserv-

ing privacy. Unlike traditional centralized learning,

which requires transferring data to a central server,

FL enables training to occur locally on client devices,

addressing critical concerns about data conﬁdential-

ity (Kairouz and et al., 2021). By maintaining data

on client devices, FL mitigates privacy risks while

facilitating the training of large-scale machine learn-

ing models. This paradigm has been signiﬁcantly

adopted in applications including predictive text input

(e.g., GBoard (Hard et al., 2019)), speech recognition

(e.g., Siri (Granqvist et al., 2020)), healthcare diag-

nostics (Rieke et al., 2020), and ﬁnance (Liu et al.,

2020; Long et al., 2020).

https://orcid.org/0000-0002-1557-4870

https://orcid.org/0000-0002-1086-1483

https://orcid.org/0000-0001-9816-7439

FL faces key challenges that limit its scalability

and practical adoption. A major issue arises from

the heterogeneity of client devices and their non-

Independent and Identically Distributed (non-IID)

data distributions. This heterogeneity leads to vari-

able update quality (Li et al., 2020; Nie et al., 2022).

Traditional aggregation methods such as Federated

Averaging (FedAVG) (McMahan et al., 2017), which

rely on weighted averaging, struggle under these con-

ditions, impairing global model generalization, con-

vergence speed, and overall performance.

Communication inefﬁciency is another critical

limitation. The standard FL pipeline transmits all

client updates to a central server indiscriminately,

leading to excessive bandwidth usage, especially

in large-scale or resource-constrained environments

such as IoT networks (Kontar et al., 2021). Tech-

niques like parameter compression (Kone

y et al.,

2017), federated dropout (Bouacida et al., 2021), and

structured updates (Zhang et al., 2024; Kone

y et al.,

2017; Wang et al., 2024) offer partial solutions but

often fail to adapt effectively to the dynamic and dis-

tributed nature of FL.

Pacioni, E., Fernández De Vega, F. and Calvaresi, D.

Towards a Meaningful Communication and Model Aggregation in Federated Learning via Genetic Programming.

DOI: 10.5220/0013380400003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1427-1431

ISBN: 978-989-758-737-5; ISSN: 2184-433X

1427

To address these limitations, this work proposes a

novel paradigm that integrates FL with Multi-Agent

Systems (MAS) modeling equipped with an adaptive

Genetic Programming (GP)-based aggregation strat-

egy.

GP has been introduced by Koza et al. (Koza,

1992) and is particularly effective in optimizing non-

linear functions and generating solutions dynami-

cally. A prominent application of GP is in solving

symbolic regression problems (Augusto and Barbosa,

2000), which involves creating mathematical models

that accurately ﬁt a given set of data points. In the

context of FL, there is a similar need for a mathe-

matical function—the aggregation function—that can

effectively handle diverse data inputs. This resem-

blance to symbolic regression makes GP a powerful

tool for evolving aggregation strategies that are cus-

tomized to accommodate heterogeneous data distri-

butions and varying client capabilities.

Therefore, we envision MAS empowering clients

(agents) to decide when and whether to communicate

with the central server, thereby optimizing bandwidth

usage and aggregation delays. Moreover, the GP-

based strategy dynamically adjusts aggregation equa-

tions to better accommodate diverse client data distri-

butions. Combining MAS for autonomy and commu-

nication efﬁciency, and GP-based aggregation for ro-

bustness, the proposed framework aims to overcome

key bottlenecks in FL, delivering scalable, resilient,

and personalized learning systems.

The remainder of this paper is structured as fol-

lows: Section 2 outlines the motivations and chal-

lenges driving this work. Section 3 details the pro-

posed paradigm, emphasizing the integration of MAS

and GP-based aggregation. Section 4 presents a

roadmap for implementation and future exploration.

Finally, Section 5 discusses the potential impact and

concludes the paper.

2 MOTIVATIONS

While FL presents a promising approach for dis-

tributed learning, its current implementations face

two critical shortcomings: communication inefﬁ-

ciency and limited adaptability in aggregation strate-

gies. These issues are particularly pronounced in non-

IID scenarios, where treating clients uniformly often

results in inefﬁciencies (Wang et al., 2020). Further-

more, transmitting all client updates indiscriminately

imposes substantial communication costs, particu-

larly in bandwidth-constrained environments. Client

selection approaches handle the process server-side,

after receiving the weights, this does not help to im-

prove efﬁciency (Fu et al., 2023). Existing approaches

to improve communication efﬁciency, including pa-

rameter compression (Kone

y et al., 2017), feder-

ated dropout (Bouacida et al., 2021), and structured

updates (Zhang et al., 2024; Kone

y et al., 2017;

Wang et al., 2024), partially mitigate these issues

by reducing the volume of transmitted data. How-

ever, these methods do not empower clients to per-

form local selection of updates, resulting in the trans-

mission of irrelevant data that is later discarded by

the server—introducing unnecessary overhead. Simi-

larly, static aggregation methods like FedAVG fail to

account for the unique characteristics of client data,

particularly in non-IID scenarios (Wang et al., 2020).

FedGR (Zeng et al., 2024) introduces an innova-

tion in FL by using a genetic algorithm for dynamic

client clustering and a relay strategy to train models

within groups. This approach stands out for its abil-

ity to reduce the impact of statistical heterogeneity

by improving the convergence of the overall model.

However, FedGR is limited to a static aggregation

strategy within each group and does not consider dy-

namic customization of aggregation algorithms.

An interesting analysis of aggregation techniques

in Federated Learning has recently been produced (Qi

et al., 2024) which highlights as major problems (i)

statistical heterogeneity of data, (ii) communication

bottlenecks, (iii) security and privacy, (iv) model cus-

tomization, and (v) client evaluation and selection.

This work addresses some of these limitations

by integrating MAS to enable autonomous decision-

making at both the client and server levels. The local

agents assess their “distance” from the global model

to determine whether to transmit updates, while the

server (aggregating agent) dynamically decides when

to distribute aggregated models or skip aggregation

rounds. Combined with a GP-based aggregation strat-

egy, this approach can enhance communication efﬁ-

ciency and ensure adaptability to evolving data distri-

butions, capturing the peculiarities of each client

3 NOVEL PARADIGM

The proposed paradigm integrates a MAS-driven

communication framework with a GP-based adaptive

aggregation strategy, effectively addressing the dual

challenges of communication efﬁciency and aggre-

gation adaptability. Figure 1 illustrates the enhanced

FL pipeline, incorporating MAS for communication

optimization and GP for dynamic aggregation.

MAS for Communication Efﬁciency. In the

proposed framework, each client operates as an

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1428

autonomous agent able to assess the relevance of

its updates based on a deﬁned distance metric w.r.t.

the global model. Agents decide independently

whether to transmit updates, thereby reducing unnec-

essary bandwidth consumption. Concurrently, the

central server employs MAS principles to evaluate

when to distribute updated global models, ensuring

synchronization is both timely and efﬁcient. This

decentralized decision-making mechanism enhances

scalability and efﬁciency, particularly in resource-

constrained environments.

GP-Based Aggregation. Conventional aggregation

techniques (i.e., FedAVG) rely on static equations that

fail to account for the heterogeneity of client data,

particularly in non-IID scenarios. The proposed GP-

based aggregation method dynamically evolves ag-

gregation equations to adapt to the client data distribu-

tions. This adaptive strategy balances local data char-

acteristics with global model performance, yielding a

more robust solution for non-IID environments. Fur-

thermore, by incorporating transfer learning within

the GP-based aggregation framework, the computa-

tional overhead of evolving new equations in each

round is signiﬁcantly minimized.

By combining these innovations, the proposed

paradigm offers a scalable solution for FL that main-

tains high performance and supports personalized

models, even in complex and heterogeneous data en-

vironments.

4 ROAD MAP

This section outlines a roadmap to address key

challenges in FL and validate the integration of MAS

and GP for scalable, efﬁcient, and decentralized

learning.

Phase 1: Developing the GP-Based Aggregation

Method – This phase focuses on creating a GP-based

technique to evolve aggregation equations tailored

to heterogeneous environments dynamically. A

key aspect is deﬁning the ﬁtness function. Ini-

tially, a single-objective function will be employed,

leveraging metrics such as accuracy, precision, or

mean Average Precision (mAP), depending on the

neural network type. Subsequently, a multi-objective

function will be introduced, incorporating factors like

convergence speed, model execution time, and model

quality to optimize performance comprehensively.

To validate the proposed new aggregation method,

a systematic comparison will be made with major

aggregation methods in the literature, including

FedAvg, Scaffold, MOON, Zeno, Per-FedAvg, Fed-

Prox, FedOpt, FedRS, and FedGR. Moreover, a test

can be carried out with FedGR in combination with a

dynamic aggregation algorithm and compared with

the current results. Comparisons will be based on key

metrics such as global model accuracy, computational

efﬁciency, robustness to non-IID data, and resilience

to malicious clients. This approach will highlight the

advantages and limitations of each method, providing

a basis for establishing practical guidelines on using

different aggregation strategies in speciﬁc scenarios.

Phase 2: Transfer Learning for GP Aggregator –

Although GP could have a signiﬁcant computational

impact, mainly due to the costs of the ﬁtness function

that they have to evaluate each individual (neural net-

work) and to do so, they have to apply the inference

process and compute metrics; we believe that the

beneﬁts of its ability to dynamically adapt to non-IID

features in the data outweigh these costs. However,

to reduce computational overhead, transfer learning

mechanisms will be integrated into the GP-based

aggregation process. This approach involves two

strategies: (i) performing a single evolution in

the ﬁrst aggregation round and reusing the same

expression tree across subsequent iterations, thereby

maximizing resource efﬁciency, and (ii) transferring

the best individual from the previous evolution to

form the new population. The latter strategy offers

two scenarios: either the transferred individual is

supplemented by randomly generated individuals, or

the entire population is derived through mutations of

the selected individual.

Phase 3: MAS for Communication Efﬁciency

– MAS-based mechanisms will enable autonomous

decision-making for clients and servers, optimizing

communication and synchronization at run time. This

phase also opens avenues for further investigation: (i)

reﬁning MAS to manage large-scale FL deployments

with thousands of clients and (ii) exploring the inte-

gration of evolutionary methods within MAS to cus-

tomize the distance threshold dynamically.

5 DISCUSSION AND

CONCLUSIONS

This paper introduces a novel paradigm for FL, com-

bining MAS-driven communication and aggregation

through GP to address challenges in non-IID environ-

ments. Enabling autonomous decision-making and

dynamic aggregation, we aim to reduce communica-

Towards a Meaningful Communication and Model Aggregation in Federated Learning via Genetic Programming

1429

Aggregation!

Agent

Aggregation Phase

FedGP

Local!

Model

Agent 1

Local!

Model

Agent 2

Local!

Model

Agent n

Global Model Propagation

Leggend

Local Model Propagation

Local Model Training

Autonomous decision-making

Figure 1: MAS integration (communication efﬁciency) and GP-based aggregation (model quality).

tion overhead, accelerate convergence, improve the

generalization, and enhance personalization.

The impact of this work is threefold: scientif-

ically, it advances FL with scalable and adaptable

mechanisms; practically, it offers robust solutions for

applications in healthcare, ﬁnance, and mobile sys-

tems; socially, it promotes improved privacy and fair-

ness in model development.

As a position paper, this work highlights the

value of hybridizing AI techniques to optimize FL

paradigms, paving the way for higher-quality solu-

tions, for greater optimization, and broader real-world

applicability.

ACKNOWLEDGEMENTS

This work was partially supported by the

HES-SO RCSO ISNet HARRISON grant

(WP2), the Spanish Ministry of Economy

and Competitiveness (PID2020-115570GB-

C21, PID2023-147409NB-C22), funded by

MCIN/AEI/10.13039/501100011033, and the

Junta de Extremadura (GR15068).

REFERENCES

Augusto, D. A. and Barbosa, H. J. C. (2000). Symbolic

regression via genetic programming. In Proceedings

of the Sixth Brazilian Symposium on Neural Networks,

volume 1, pages 173–178, Rio de Janeiro, Brazil.

Bouacida, N., Hou, J., Zang, H., and Liu, X. (2021). Adap-

tive federated dropout: Improving communication ef-

ﬁciency and generalization for federated learning. In

IEEE INFOCOM 2021 - IEEE Conference on Com-

puter Communications Workshops (INFOCOM WK-

SHPS), pages 1–6.

Fu, L., Zhang, H., Gao, G., Zhang, M., and Liu, X.

(2023). Client selection in federated learning: Prin-

ciples, challenges, and opportunities. IEEE Internet

of Things Journal, 10(24):21811–21819.

Granqvist, F., Seigel, M., van Dalen, R., Cahill, A., Shum,

S., and Paulik, M. (2020). Improving on-device

speaker veriﬁcation using federated learning with pri-

vacy. arXiv preprint.

Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays,

F., Augenstein, S., Eichner, H., Kiddon, C., and Ram-

age, D. (2019). Federated learning for mobile key-

board prediction. arXiv preprint.

Kairouz, P. and et al. (2021). Advances and open problems

in federated learning. Now Foundations and Trends.

Kone

y, J., McMahan, H. B., Yu, F. X., Richt

arik, P.,

Suresh, A. T., and Bacon, D. (2017). Federated learn-

ing: Strategies for improving communication efﬁ-

ciency.

Kontar, R., Shi, N., Yue, X., Chung, S., Byon, E., Chowd-

hury, M., Jin, J., Kontar, W., Masoud, N., Nouiehed,

M., Okwudire, C. E., Raskutti, G., Saigal, R., Singh,

K., and Ye, Z.-S. (2021). The internet of federated

things (ioft). IEEE Access, 9:156071–156113.

Koza, J. R. (1992). Genetic Programming: On the Pro-

gramming of Computers by Means of Natural Selec-

tion. MIT Press.

Li, X., Huang, K., Yang, W., Wang, S., and Zhang, Z.

(2020). On the convergence of fedavg on non-iid data.

arXiv preprint.

Liu, Y., Ai, Z., Sun, S., Zhang, S., Liu, Z., and Yu, H.

(2020). Fedcoin: A peer-to-peer payment system for

federated learning. In Yang, Q., Fan, L., and Yu, H.,

editors, Federated Learning: Privacy and Incentive,

pages 125–138, Cham. Springer International Pub-

lishing.

Long, G., Tan, Y., Jiang, J., and Zhang, C. (2020). Federated

learning for open banking. In Yang, Q., Fan, L., and

Yu, H., editors, Federated Learning: Privacy and In-

centive, pages 240–254, Cham. Springer International

Publishing.

McMahan, H. B., Moore, E., Ramage, D., Hampson, S., and

uera y Arcas, B. (2017). Communication-efﬁcient

learning of deep networks from decentralized data. In

Proceedings of the 20th International Conference on

Artiﬁcial Intelligence and Statistics (AISTATS).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1430

Nie, W., Yu, L., and Jia, Z. (2022). Research on aggregation

strategy of federated learning parameters under non-

independent and identically distributed conditions. In

Proceedings of the 2022 4th International Conference

on Applied Machine Learning (ICAML), pages 41–48.

Qi, P., Chiaro, D., Guzzo, A., Ianni, M., Fortino, G., and

Piccialli, F. (2024). Model aggregation techniques in

federated learning: A comprehensive survey. Future

Generation Computer Systems, 150:272–293.

Rieke, N., Hancox, J., Li, W., Milletar

ı, F., Roth, H. R.,

Albarqouni, S., et al. (2020). The future of digital

health with federated learning. npj Digital Medicine,

3(1):119.

Wang, J., Liu, Q., Liang, H., Joshi, G., and Poor, H. V.

(2020). Tackling the objective inconsistency prob-

lem in heterogeneous federated optimization. In Pro-

ceedings of the 34th International Conference on Neu-

ral Information Processing Systems, NIPS ’20, Red

Hook, NY, USA. Curran Associates Inc.

Wang, L., Li, J., Chen, W., Wu, Q., and Ding, M.

(2024). Communication-efﬁcient model aggregation

with layer divergence feedback in federated learning.

IEEE Communications Letters, 28(10):2293–2297.

Zeng, Y., Zhao, K., Yu, F., Zeng, B., Pang, Z., and Wang, L.

(2024). Fedgr: Genetic algorithm and relay strategy

based federated learning. In Proceedings of the 27th

International Conference on Computer Supported Co-

operative Work in Design (CSCWD).

Zhang, W., Zhou, T., Lu, Q., Yuan, Y., Tolba, A., and Said,

W. (2024). Fedsl: A communication-efﬁcient feder-

ated learning with split layer aggregation. IEEE Inter-

net of Things Journal, 11(9):15587–15601.

Towards a Meaningful Communication and Model Aggregation in Federated Learning via Genetic Programming

1431