| Popis: |
Vehicular traffic congestion is a major problem all over the world with significant economic and environmental impact. Adapting traffic signal durations is one method to alleviate this problem and has the advantage that it does not require any significant changes to existing infrastructure such as traffic poles and roads. Reinforcement learning is suitable for adapting the traffic signal durations since it does not require any prior knowledge of traffic patterns, which are time-varying and a priori unknown. In reinforcement learning based schemes proposed in previous studies, the algorithm for adapting the durations of a traffic signal takes into account only the vehicle queue lengths at that signal. In this paper, we propose a novel Q-Learning based algorithm, which adapts the durations of a signal by taking into account the vehicle queue lengths at all the signals that are n or fewer hops away from the signal, where n is a parameter that enables us to trade performance with computational complexity. In particular, our simulations show that as n increases, the performance of the algorithm improves in terms of travelling time, number of moving vehicles as well as Carbon dioxide emissions, albeit at the expense of an increase in computation time. Further, we consider the case in which n is increased gradually as time progresses, and show that the performance achieved is significantly better than in the case where n is a constant. |
