New method uses old bus data to prevent buses from bunching up

Researchers developed a way to train bus control policies using only historical GPS data.

Buses often bunch up even when they start at equal headways. This problem makes passenger waiting times longer and creates load imbalances. It also hurts service reliability.

A new method called Offline Graph-based Multi-agent Actor-Critic (Offline-GMAC) manages bus holding. It uses logged GPS data from normal uncontrolled operations. This avoids disrupting the bus system during training.

The method uses general graph neural networks for multi-agent credit assignment. This is a way to assign credit among multiple agents. It also uses a data augmentation scheme. This scheme creates unperformed holding actions from historical trajectories.

The scheme uses the equivalence between stochastic link travel time delays and deliberate holding interventions. The 1-Wasserstein distance between synthesized and target state transition distributions is bounded. This bound scales with the standard deviation of link travel times plus a tunable synthesis-dependent term.

Evaluations using real-world bus data show this offline method works well. It achieves performance comparable to, and sometimes better than, its online counterpart.

Key facts

  • Offline-GMAC trains bus holding policies using only historical GPS data.
  • The method uses a data augmentation scheme to synthesize unperformed actions.
  • The approach avoids the need for disruptive online interaction or exploration.

All sections

Edition