Towards AI Agents for Selecting Architectural Patterns in Federated Learning Systems

Recent studies have shown that the adoption of architectural patterns in Federated Learning systems can lead to significant advantages in terms of system efficiency (e.g., improved model accuracy and faster training time). However, identifying which architectural patterns or combinations of them can lead to improvements in a given system remains a manual and error-prone process. To address this challenge, we take an initial step toward enabling the automated (de)selection of architectural patterns in Federated Learning systems. We extend our benchmarking platform, namely AP4Fed, with the conceptualization of AI Agents that are in charge of reasoning about possible architectural alternatives. At the beginning of each Federated Learning round, one or multiple AI agents analyze current system metrics (e.g., model accuracy, total round time) and contextual information (e.g., resource capabilities of participating clients), collaboratively deciding which architectural patterns should be (de)activated. This approach can support software architects by introducing an AI-driven decision-making mechanism that dynamically selects architectural patterns during Federated Learning simulations, aiming to improve the system efficiency and to reduce the manual tuning effort.