Chuyển tới nội dung chính

Introduction to GreenWave AI

GreenWave AI is a smart traffic light control system utilizing Artificial Intelligence (AI) and Reinforcement Learning, completely integrated with the FIWARE platform and SUMO traffic simulation.

Overview​

Unlike traditional traffic light systems that use fixed-time counters, GreenWave AI analyzes traffic flow in real-time to make the most optimal control decisions.

The system operates based on the principle: "Green lights are only for directions that truly need them".

Key Features​

  • Independent Control: Each traffic light pillar is analyzed and controlled independently by AI, not rigidly dependent on other lights.
  • Real-time: Decisions are made every 2 seconds based on current data from sensors.
  • Conflict-Free: The algorithm ensures safety, ensuring that signal conflicts never occur.
  • Multi-Objective: Simultaneously optimizes multiple metrics: queue length, waiting time, and vehicle density.

System Architecture​

GreenWave AI operates according to the following closed-loop pipeline:

Ai-Architecture.png

  1. Data Collection: Sensors (Induction loops) in SUMO collect information about speed and vehicle count.
  2. Context Broker: Data is normalized into TrafficFlowObserved and sent to Orion-LD.
  3. AI Processing: The AI Agent receives notifications from Orion, analyzes the state, and makes a phase decision.
  4. Execution: Control commands are sent back via the IoT Agent to change the light status in the simulation.

Artificial Intelligence & Algorithms​

GreenWave employs Reinforcement Learning (RL) to solve the Traffic Signal Control (TSC) problem. Unlike traditional timer-based systems or static sensors, this AI learns optimal strategies by interacting with the simulation.

Algorithm: Deep Q-Network (DQN)​

We use a Double DQN (Deep Q-Network) approach to stabilize learning.

  • Model Architecture: A multi-layer perceptron (Neural Network) built with TensorFlow/Keras.
    • Input Layer: 4 Neurons (State State).
    • Hidden Layer 1: 128 Neurons (ReLU activation, Dropout 0.2).
    • Hidden Layer 2: 128 Neurons (ReLU activation, Dropout 0.2).
    • Hidden Layer 3: 64 Neurons (ReLU activation).
    • Output Layer: 2 Neurons (Linear activation) representing Q-values for each action.
  • Optimizer: Adam (Learning Rate = 0.0005).
  • Loss Function: Mean Squared Error (MSE).

The "Dataset" (Experience Replay)​

In Reinforcement Learning, there is no static "dataset" (like CSVs or Images). The AI creates its own dataset to learn from by interacting with the environment:

  • Source: Analysis of real-time simulation frames (SUMO).
  • Experience Replay Buffer: Stores the last 10,000 interaction steps.
    • Format: (State, Action, Reward, Next_State).
  • Training: The model samples a random "batch" (size 64) from this buffer to train itself, ensuring it learns from both recent and past experiences (preventing forgetting).

State Space (Input)​

The AI "sees" the intersection through a vector of 4 values:

  1. Queue Length (Detector 1): Number of cars waiting at the North/South arm.
  2. Queue Length (Detector 2): Number of cars waiting at the East/West arm.
  3. Current Phase: Which light is currently green? (Index).
  4. PM2.5 Emission: Total air pollution calculated from vehicle emissions in the area.

Action Space (Output)​

The AI can make 2 Decisions at every step:

  • Action 0 (HOLD): Keep the current light Green.
  • Action 1 (SWITCH): Switch the light to Red (and the other to Green).
    • Constraint: The system enforces a Minimum Green Time (e.g., 10 seconds) to prevent chaotic, rapid switching.

Goal & Reward Function​

The AI's goal is to maximize the "Reward". Our reward function is designed to balance traffic flow and environmental impact:

Reward = (0.6 × -TotalQueue) + (0.4 × -TotalPM2.5)

  • Logic: The reward is always negative (a penalty). The AI tries to get this number as close to zero as possible.
  • Traffic Priority (60%): Minimizing queues is the main goal.
  • Eco Priority (40%): If two actions reduce queues equally, the AI picks the one that generates less pollution (e.g., keeping heavy trucks moving).

Performance & Effectiveness​

  • Adaptive: Unlike fixed-time lights, this system adapts to rush hour vs. midnight traffic automatically.
  • Multi-objective: It solves not just for speed, but for Sustainability (Green Wave).
  • Evaluation: During training, we observed a reduction in average waiting times compared to static timing, specifically in high-load scenarios (saturated intersections).