Skip to content

Key Features (Python Inference Service)

Feature List

Supported Junctions

The service manages these 5 junctions with the following action spaces:

  1. joinedS_265580996_300839357 - 4 valid phases (0, 1, 2, 3)
  2. 300839359 - 2 valid phases (0, 1)
  3. 265580972 - 2 valid phases (0, 1)
  4. 1270712555 - 2 valid phases (0, 1)
  5. 8541180897 - 2 valid phases (0, 1)

Neural Network Architecture

The model uses a shared RNNAgent with the following structure:

  • Input Layer: Concatenates observations + agent ID one-hot encoding
  • Observation vector: up to 19 floats (zero-padded to this size)
  • Agent ID one-hot: 5 floats (one per junction)

  • Total input: 24 floats

  • Hidden Layer: GRUCell

  • Dimension: 128

  • Maintains stateful hidden state per junction across calls

  • Output Layer: Fully connected

  • Outputs logits for all possible actions (4 max)

  • Masked by junction-specific action availability

  • Activation: ReLU on first layer, softmax on output

Observation Format

Each observation contains traffic state information including: - Phase encoding (one-hot representation of current signal phase) - Minimum green flag (whether minimum green time has elapsed) - Lane queue lengths for approach lanes - Vehicle wait times - Gap detection values

Example observation sizes: - joinedS junction: 19 floats (full-featured) - Other junctions: ~8-10 floats (zero-padded to 19 internally)