Winning project of BITEhack 2023 hackathon in AI category.
This is a Reinforcement Learning system that trains a traffic light agent to optimally switch based on real time traffic observations.
The optimization goal is the minimization of wasted fuel - fuel used while stuck in traffic.
Each traffic light sees cars in the entire city, other traffic light states and its own position.
The PPO (Proximal Policy Optimization) agent has been trained in such an environment.
Results have been compared with a baseline model (switching lights every X ticks). The learned PPO agent outperforms the baseline.
Baseline agent on the left, trained PPO on the right.
The reward is fuel use efficiency (100% means no fuel was wasted while not moving).
Total computational time was 6 h.
- install python 3.10 and pytorch
pip install -r requirements.txt
With enough RAM, VRAM, CPU cores and a good GPU run:
python train.py
A pretrained checkpoint is provided in checkpoints/.
Specify the trained checkpoint path in simulation.py and run:
python simulation.py
The step function takes an action mapping of traffic light ids to their requested states.
Each traffic light has a switching frequency limit. This is to stop rapid state changes.
It returns the environment observation and reward. The environment never terminates by itself.
The observation is designed for a convolutional neural network. It contains 6 channels, each NxM (size of grid):
- the map (1 where road)
- car positions (1 where car)
- this agent position (1 where this agent junction)
- 1 where lights are in state 0
- 1 where lights are in state 1
- all available junctions
To speed up training, the map was removed from the observation and the CNN was replaced with a simple FCN.
The reward in each step is calculated as (number of cars that moved / total cars).
The reward is summed over all agents and environment steps.
For simulation evaluation, an average of 100 step rewards is taken.
An amazing team from AGH UST Computer Science students: