Learning and managing stochastic network traffic dynamics: an iterative and interactive approach

ABSTRACT

This study examines the potential of an iterative and interactive approach to learn network traffic dynamics and optimise tolling strategies considering time-varying stochastic traffic. A tractable ‘truth model’ based on the stochastic Macroscopic Fundamental Diagram is developed to represent the transportation system to be learned and managed. A ‘twin model’ that mirrors the truth model is formulated and calibrated for testing and optimising tolling adjustment strategies with the help of reinforcement learning. The optimised prices are then put into the ‘truth model’ to evaluate network efficiency improvement. The above procedure is iterative and interactive, which can be applied for congestion management in the period-to-period tolling adjustment fashion. Numerical studies show that the proposed iterative and interactive pricing strategy is able to enhance network efficiency even under limited information and/or inaccurate learning of the system. This illustrates the great potential of utilising iterative and interactive frameworks for congestion management.

KEYWORDS:

• Within-day traffic dynamics
• day-to-day
• iterative and interactive
• dynamic pricing
• stochastic MFD

Acknowledgments

The authors would like to thank the anonymous referees for their detailed and constructive comments, which have helped improve the manuscript substantially.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 As reported in the study of Ma et al. (2021) based on real data from Sydney Public Transport System, ρ is around 0.83, i.e. $83\mathrm{\%}$ travelers will stick with their previous travel choices, while $17\mathrm{\%}$ will reconsider. $\rho =0.83$ will be used in the numerical study.

2 The average speed ${\overline{v}}_{i,m}^w(t,q)$ in region i (average for the trip distance within region i) for traveler choosing route m with O-D pair w at the departure time t on day q can be calculated as follows: ${\overline{v}}_{i,m}^w(t,q)=\frac{l_i^w}{T_i^w(t,q)}$, where $l_i^w$ is the traveler's travel distance in region i and $T_i^w(t,q)$ is his or her experienced travel time in region i.

3 https://pygad.readthedocs.io/en/latest/.