Ramp metering is a freeway control method capable of preventing operational breakdown due to excessive total demand by limiting entry to the main line using on-ramp traffic lights. When the control strategy is properly designed, it has the potential to eliminate congestion near on-ramps, avoid blockage of exit ramps, influence route choice behaviour, and enhance traffic safety via safer merging. None of the existing ramp metering strategies is adaptive from a control theory point of view, i.e. controller parameters are not changing while working.; This thesis introduces a new reinforcement learning approach to closed-loop adaptive optimal freeway ramp metering. Reinforcement learning is a powerful machine-learning tool used for unsupervised learning especially in stochastic environments. Its continuous learning capability enables a truly adaptive control strategy. |