Resumo:
The number of vehicles on the streets across the world has quickly grown in the
last decade, directly impacting how urban traffic is managed. The signalized junctions
control is a vastly known and studied problem. Although an increasing number of technologies
is explored and used to solve it, there still are challenges and opportunities to deal
with it, especially when considering the inefficiency of the widely known fixed time traffic
controllers, which are incapable of dealing with dynamic events. This study aims to apply
Hierarchical Reinforcement Learning (HRL) on the control of a signalized vehicular junction
and compare its performance with a fixed time traffic controller, configured using the
Webster Method. HRL is a Reinforcement Learning (RL) variation, where secondary objectives,
represented by sub-policies, are organized and proposed in a hierarchical model,
managed by a macro-policy, responsible for selecting said sub-policies when those are
capable of reaching its best results, where The Q-Learning Framework rules both sub
and macro policies. Hierarchical Reinforcement Learning was chosen because it combines
the ability to learn and make decisions while taking observations from the environment,
in real-time, a typical ability from Reinforcement Learning, with a Divide to Conquer
approach, where the problem is divided into sub-problems. These capabilities bring to a
highly dynamic problem a more significant power of adaptability, which is impossible to
be taken into account when using deterministic models like the Webster Method. The
test scenarios, composed of several vehicle fluxes applied to a cross of two lanes, were
built using the SUMO simulation tool. HRL, its sub-policies and the Webster Method
are applied and assessed through these scenarios. According to the obtained results, HRL
shows better results than the Webster Method and its isolated sub-policies, indicating a
simple and efficient alternative.