Constrained Bayesian Optimization for Hierarchical Federated Learning in IoT Networks for Plant Disease Classification

A sample-efficient framework for jointly tuning deep learning backbones, aggregation strategies, and communication rounds under energy, latency, and accuracy constraints

Athanasios Papanikolaou; Athanasios Tziouvaras; Apostolos Xenakis; Periklis Chatzimisios; Shameem A. Puthiya Parambath; George Floros; Enrica Zereik; Ivan Petrovic; Fabio Bonsignorio· 2026· DOI 10.48550/arXiv.2609.06830

The core problem

Deploying Hierarchical Federated Learning (HFL) in resource-constrained Internet of Things (IoT) environments demands careful configuration to balance predictive performance against energy consumption and execution time. This challenge is especially acute in smart agriculture, where distributed IoT devices must support automated plant disease classification under limited computational and communication budgets. The authors address this by formulating the HFL configuration problem as a constrained optimization task and solving it with Bayesian Optimization (BO). The framework jointly explores three key decision variables: the deep learning backbone architecture, the federated aggregation strategy, and the number of communication rounds. The federation size is not treated as a free variable; instead, it is determined by the spatial coverage requirements of the agricultural deployment. A weighted objective function encodes user-defined trade-offs among energy consumption, execution time, and predictive performance, while explicit constraints enforce deployment-specific resource and accuracy requirements. The paper evaluates the approach on an IoT-based plant disease classification tas

Innovation

Experimental results across 30 independent optimization runs demonstrate that the constrained BO framework explores only 11.11% of the full search space while consistently identifying solutions within 1% of the exhaustive-search optimum. The mean optimality gap is merely 0.056%, indicating near-optimal configuration selection with substantially reduced computational effort. The framework successfully handles the trade-offs among energy consumption, execution time, and predictive performance, returning configurations that satisfy all deployment-specific constraints. The evaluation considered multiple deep learning architectures, federated aggregation strategies, and communication-round settings, confirming the generality of the approach across the tested HFL design space. The low optimality gap and high sample efficiency highlight the practical viability of constrained BO for tuning HFL deployments in resource-constrained IoT environments.

Deploying Hierarchical Federated Learning (HFL) in resource-constrained Internet of Things (IoT) environments demands careful configuration to balance predictive performance against energy consumption and execution time. This challenge is especially acute in smart agriculture, where distributed IoT devices must support automated plant disease classification under limited computational and communication budgets. The authors address this by formulating the HFL configuration problem as a constrained optimization task and solving it with Bayesian Optimization (BO). The framework jointly explores three key decision variables: the deep learning backbone architecture, the federated aggregation strategy, and the number of communication rounds. The federation size is not treated as a free variable; instead, it is determined by the spatial coverage requirements of the agricultural deployment. A weighted objective function encodes user-defined trade-offs among energy consumption, execution time, and predictive performance, while explicit constraints enforce deployment-specific resource and accuracy requirements. The paper evaluates the approach on an IoT-based plant disease classification task using multiple deep learning architectures, aggregation strategies, and communication-round settings.

The proposed framework casts HFL configuration as a constrained black-box optimization problem. Let x = (a, s, r) denote the configuration vector, where a is the deep learning backbone architecture, s is the federated aggregation strategy, and r is the number of communication rounds. The federation size N is fixed a priori based on the spatial coverage required by the agricultural deployment. The objective is to minimize a weighted cost function:

Why it matters

The results underscore the effectiveness of constrained Bayesian Optimization for HFL configuration in IoT-based smart agriculture. By exploring only 11.11% of the search space, the method drastically reduces the computational burden associated with exhaustive search, which is critical when each evaluation involves training and simulating a federated learning deployment. The mean optimality gap of 0.056% confirms that the surrogate model and acquisition function guide the search toward near-optimal solutions reliably. The explicit handling of constraints ensures that returned configurations are not only performant but also compliant with real-world resource and accuracy requirements. The joint optimization of backbone architecture, aggregation strategy, and communication rounds captures important interactions that are often overlooked when these parameters are tuned independently. A limitation is that federation size is fixed by spatial coverage rather than optimized, which may leave some performance on the table in deployments where coverage requirements are flexible. Future work could extend the framework to dynamic federation sizes, online adaptation to changing conditions, and integration with more diverse IoT hardware. Overall, the study provides a practical and sample-efficient methodology for deploying HFL in precision agriculture, with potential transferability to other resource-constrained federated learning applications.

Who should read this

CS practitioners and researchers

Opening member content…