An Ensemble Method for Traffic Light Management

The main tool used to conduct experiments is SUMO-RL framework.

Ensemble methods used:

• Majority Voting
• Soft Voting
• Transformed Rank Voting
• Average Voting - Boltzmann Probs

Reward is defined as the change of the cumulative vehicle delay $r_{t} = D_{a_{t+1}} - D_{a_{t}}$

Action: Choose the next Green phase

State Representation: Vector of dimension $R^{GreenPhases+2*Lanes}$

Environment: Simulation of Urban MObility (SUMO)

Our Approach

A single intersection consisted of:

• 2 Incoming - 2 Outgoing approaches
• Totally 8 Lanes
• 8 Permitted Movement Signals
• Sythetic Data built on SUMO:
  • Approach Length: $300m$
  • Cycle Traffic Plan Duration: $82s$