Intelligent Traffic Signal Control

This Project is a traffic control system based on DQN (arxiv:1312.5602) on Vissim. It's an original implement that intelligent traffic signal control via deep reinforment learning on partial urban traffic net. Choosing fine hyper-parameters, agent could learn to how to improve the performance of global net in a long term.

Dependencies

• Vissim 4.3.0
• Python 3.5
• Tensorflow 1.2.0
• other common packages like pandas numpy matplotlib pywin32
• ...

About Vissim

VisEnv.py wrapped the orignal api into the open.ai style. For now, speed, travel time, queued vehicles count interfaces are provided. Use this like:

fron vis_env import * 
env = VisEnv()
...
for epi in range(episodes):
    env.reset()
    env.test = True
    for _ in range(steps):
        next_state, reward, done = env.step(action)
    env.write_summary(epi, dir)

Experiments

The performance of DQN is not so good among the series of reinforcement learning algorithm, but agent are still capble to act appropriately in our traffic enviroment.

TODO

More reinforment learning models like dueling-DQN, DDPG, to further improve the performance of agent, and to solve the large discrete actions space problem. Intelligent Traffic Signal Control