Multi-agent Architecture for Federated Learning

The concept of federated learning has become widespread in working with data, mainly due to the fact that it allows training on data directly on the nodes where they are stored. As a result, no data transfer is required. After the training is completed on each node, only the trained model is transmitted to the central server for aggregation. Multi-agent systems behave in a similar way, because agents allow you to train machine learning models on local devices, while preserving confidential information. The ability of agents to interact with each other makes it possible to generalize (aggregate) such models and reuse them. This article presents the architecture of multi-agent systems for federated learning. It highlights the elements that make up the agent platform and the structure of the JADE platform. Describes the lifecycle of all agents used to perform a full training cycle in the MAC\_FL environment. The configurations of agent placement for each of the proposed architectures of multi-agent systems of federated learning are analyzed and described: centralized, decentralized and hierarchical.