Distributed learning for multi-agent control of a dynamic system

This thesis describes an investigation of self-organising, distributed control of dynamic, nonlinear systems. The distribution is achieved through a multi-agent based approach. The self-organisation is addressed through reinforcement learning. The feasibility is tested using a well-established agent framework: JADE. The target system for the study is a simulation of a well-known laboratory demonstrator, the twin-rotor MIMO system, but configured to introduce strong cross-couplings in its non-linear dynamics. A multi-agent PID controller is developed as an interim solution to test the feasibility of the use of JADE for control purposes. An overarching constraint on the development of any solutions was that the plant knowledge was minimal, which placed great importance on the need for a self-organising scheme. Results of a developed system are presented against the context of more conventional control methodologies.