Multi-Agent Model Predictive Control Cooperation in a Real Eight-Tank Plant based on Reinforcement Learning

In this work, a negotiation agent trained by deep reinforcement learning methodology is employed in order to achieve control objectives in a multi-agent cooperative distributed implementation of model-based predictive control (MPC). The negotiation agent is located in the upper layer of a hierarchical control architecture to achieve a consensus between the different local MPC controllers, providing weighting coefficients for the negotiation process between the control sequences provided by the local controllers. The training algorithm for the reinforcement learning of this agent is the deep deterministic policy gradient algorithm (DDPG). The validation of the negotiation agent has been performed successfully both in simulation and in a real laboratory plant composed of eight coupled water tanks, which is a non-linear system characterized by high interaction between subsystems.