Abstract:
Traffic demand increases day by day and leads to traffic congestion problem in big
cities. One solution is to build new traffic infrastructure that might cause a huge burden on
country’s economy. Another solution is to optimize existing traffic network by controlling
traffic lights at intersections. The advancement in Reinforcement Learning technique has
shown a potential to solve complex nature of traffic congestion problem. In this thesis
research, reinforcement learning was implemented in context of traffic control system to
investigate the improvement made in traffic flow at intersection. The two reinforcement
learning agents, turn-based and time-based, were designed to optimize traffic flow at
intersection. Turn-based agent opens the traffic signal for that side of intersection that has
high traffic flow, in contrast to this time-based agent follows the fixed phase cycle and
changes the phase duration on the basis of traffic. In both turn-based and time-based agents,
the state is obtained by encoding scalar number of queue length. The agents were trained
using deep Q-learning approach. In order to improve the agent’s learning experience four
different traffic scenarios were generated. The simulation software SUMO was used to
generate simulations for agent training and evaluation. The result shows that turn-based and
time-based agents perform better than conventional traffic light control system in all traffic
scenarios. Both agents are designed according to the need of practical application and
experiment is conducted to evaluate them in simulation on real-world map of Islamabad city.
This experiment includes four intersection on Kashmir Highway form G-11 to G-9. The
result shows that reinforcement learning agents perform better than conventional light traffic
control system at all four intersections
