This paper introduces the Multi-Agentic Architecture Specialised for Control (MAASC). This framework transforms general-purpose AI models into domain-specific controllers by arranging multiple specialised agents in a coordinated architecture. In MAASC, each agent is a specialisation of a general-purpose model trained to perform a distinct regulatory function, embedding control-theoretic inductive biases to guide learning. This paper demonstrates the concept through an advanced regulatory PID control (ARC) informed MAASC implementation, where neural specialisation units internalise key elements, such as gain calculation, selector logic, and split-range actuation. Within this framework, the Neuro-Controller Simple Internal Model Control (NCSIMC) is encapsulated as a specialised instance of a general-purpose ML model trained exclusively on open-loop data, in line with classical regulatory control practices. This encapsulation enables the architecture to reproduce expert-designed behaviours without requiring closed-loop training. The proposed MAASC implementation is validated through an oil artificial lift system operated by an electric submersible pump. Results show that this multi-agentic specialisation achieves regulatory performance and input constraint handling using only open-loop data, highlighting MAASC as a scalable paradigm for embedding domain expertise into AI-based control. • Using hybrid agents for PID-based control. • Building custom neural networks for control. • Inductive bias for specialising advanced control agents.
Costa et al. (Thu,) studied this question.