Effects of Social Interactions in Self-Organising Railway Traffic Management

Technical Summary: Effects of Social Interactions in Self-Organising Railway Traffic Management

Problem Statement
Real-time railway traffic management traditionally relies on centralised dispatching, which faces significant scalability challenges in dense networks, particularly regarding computation times and the integration of heterogeneous or commercially sensitive data from competing operators. While self-organising traffic management systems (SO-TMS) offer a decentralised alternative by delegating decision authority to local agents (trains), the structural parameters governing these systems require rigorous evaluation. Specifically, this paper investigates the predictive neighbourhood horizon ($T_h$), a critical design variable that determines the temporal window over which trains predict potential conflicts and establish their local interaction topology. The core problem is to understand the trade-off between local tractability (computational responsiveness) and global schedule coherence (feasibility and optimality) as $T_h$ varies.

Methodology
The study employs a closed-loop simulation framework integrating a modular SO-TMS architecture with the OpenTrack microscopic railway simulator. The SO-TMS pipeline consists of five stages:

1. Neighbourhood Identification: Trains identify potential conflicts within a time horizon $T_h$ to define a local interaction graph.
2. Hypothesis Generation: Trains solve local Mixed Integer Linear Programming (MILP) problems to generate candidate schedules (hypotheses) based on private objective functions.
3. Hypothesis Compatibility: Trains evaluate pairwise compatibility of hypotheses with neighbours.
4. Hypothesis Selection: A distributed negotiation mechanism (inspired by voter models) is used to reach a local consensus.
5. Merge: A central unit aggregates selected hypotheses into a global plan, resolving residual incompatibilities.

Experiments were conducted on a 60 km mixed-traffic corridor in Italy (Segrate to Ospitaletto) with approximately 150 daily trains. Ten perturbation scenarios were generated by sampling entrance delays from historical distributions. The system was evaluated across a range of $T_h$ values (from 300 to 1800 seconds) to measure impacts on:

• Topological metrics: Graph connectivity, component size, and node degree.
• Computational metrics: MILP optimality gaps and hypothesis generation times.
• Operational metrics: Total accumulated delay, consensus convergence speed, and the frequency of central repair interventions.

Key Results
The empirical analysis yields several counter-intuitive findings regarding the choice of $T_h$:

• Global Performance: The total network delay showed no systematic improvement with larger horizons. Across all configurations, the SO-TMS achieved an average improvement of approximately 5% over centralised control, with no significant variance attributable to $T_h$.
• Computational Burden: Increasing $T_h$ significantly increased the complexity of local MILP problems. Larger horizons led to higher optimality gaps (distance from the optimal solution) and longer computation times. Notably, at $T_h = 1500$ seconds, the decentralised system's maximum computation time (188.52s) exceeded that of the centralised solver (182.18s).
• Consensus Dynamics: Larger horizons increased the frequency of active negotiations and the maximum number of iterations required to reach consensus, with convergence times reaching up to $10^3$ steps for high $T_h$ values.
• Global Coherence: Contrary to the expectation that longer horizons improve global planning, they actually increased the frequency of incompatibilities detected during the central merge phase. This is attributed to a mismatch where local optimisations conflict with the unmodified schedules of isolated trains or those outside the immediate negotiation scope.

Significance and Claims
The paper claims that, contrary to intuition, short time horizons are entirely sufficient for effective decentralised railway traffic management. The study demonstrates that extending the predictive window compromises computational responsiveness and local solution quality without providing tangible gains in global schedule optimality.

The authors argue that the potential "myopic" nature of short-horizon planning is mitigated by the closed-loop nature of the system, where downstream replanning can easily correct local decisions. Therefore, the optimal structural condition for this SO-TMS architecture prioritises local tractability and responsiveness over extended foresight, suggesting that shorter horizons balance the need for global coherence with the practical constraints of real-time, safety-critical environments.