The rapid growth of technology and urban populations has increased the complexity of critical systems like electrical grids, communication networks, and water distribution systems. These systems, involving multiple decision-makers, face challenges such as fluctuating demand, faults, and seasonal variations, making optimal control difficult. Traditional model-based methods struggle to keep up with the evolving complexity and scale of these systems, often becoming outdated, while model-free learning-based control offers real-time adaptability without detailed system models.

This dissertation investigates learning-based adaptive controllers designed to handle interactions with other controllers, achieve multiple objectives, and remain robust to uncertainties. The central hypothesis is that optimal control for large, complex systems can be broken down into a series of decentralized optimization problems, solved independently by each controller at each time step. A game-theoretic approach, combined with reinforcement learning, models interactions among controllers as a non-cooperative game. The practical application of these decentralized methods is tested in a water distribution network with uncertain demands, implemented in Aalborg University's Smart Water Lab.