Reservoir-Based State Estimation of Partially Observable Multi-Agent Systems

• Collaborators: Francesco Sorrentino
• September 2025 - September 2026

Multi-agent complex systems often exhibit strongly coupled dynamics while only partial observations are available from individual agents. In this research, reservoir computing is explored as a data-driven approach for reconstructing the unobserved state of one agent using observations from other agents, without requiring explicit knowledge of the underlying system dynamics.

Developed an attention-based reservoir computing framework for state reconstruction in partially observable multi-agent networks. The proposed approach learns informative cross-agent interactions and selectively incorporates them into the reservoir feature space, improving reconstruction accuracy while reducing the number of required features compared with conventional reservoir computing approaches.