A formal model of multiagent systems for federated learning

Recently, the concept of federated learning has been actively developing. This is due to the tightening of legislation in the field of working with personal data. Federated learning involves performing data training directly on the nodes where the data is stored. As a result, there is no need to transfer data an-ywhere, and they remain with the owners. To generalize the trained models, they are sent to the server that performs the aggregation. The concept of federated learning is very close to a multi-agent system, since agents allow training machine learning models on local devices while maintaining confidential information. The ability of agents to interact with each other makes it possible to generalize (aggregate) such models and reuse them. Taking into account the tasks that are solved by the federated learning methods, there are several learning strategies. Learning be carried out as follows: sequentially when the model is trained in turn at each node; centrally when models are trained in parallel at each node and aggregated on a central server; or decentralized where training and aggregation is performed on each of the nodes. Interaction and coordination of agents should be carried out taking into account these learning strategies. This article presents a formal model of multi-agent systems for federated learning. It highlights the main types of agents required to complete the full cycle of federated learning: an agent that accepts a task from a user; an agent that collects information about the environment; an agent performing train-ing planning; an agent performing training on a data node; an agent providing information and access to data; an agent performing model aggregation. For each of them, the paper defines the main actions and types of messages exchanged by such agents. It also analyzes and describes the configurations of agent placement for each of the federated learning strategies.