A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems

arXiv:2606.14582 · cs.AI · Submitted 2026-06-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems".

Jane: The paper was written by Pollob Chandra Ray, Sabah Binte Noor and Fazlul Hasan Siddiqui from Department of Computer Science and Engineering, Dhaka University of Engineering & Technology.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Alright, welcome back to the show, everyone. Today we're looking at a paper that's got a mouthful of a title: "A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems." Jane, I'm going to need you to unpack that one for me.

Jane: Happy to, Tom. So the key phrase in there is "temporal planning." That's a fancy way of saying the system doesn't just figure out what actions to take, but exactly when to take them. It's like the difference between a to-do list and a detailed schedule with timestamps on every single task.

Tom: And the "heterogeneous" part? That's what caught my eye. What does that mean in the railway world?

Jane: It means not all trains and tracks are the same. In many countries, you've got different track gauges — the distance between the rails. A meter gauge train physically cannot run on a broad gauge track. The paper mentions Bangladesh and Spain as real examples where this is a daily operational headache.

Tom: So it's not just about moving trains from A to B, it's about moving them on tracks that actually fit their wheels. That's a constraint most scheduling research just ignores, right?

Jane: Exactly. And that's what makes this paper interesting. Most existing work assumes a uniform network, which is fine in theory but falls apart in practice. This team from Dhaka University of Engineering and Technology built a framework that treats gauge compatibility as a first-class constraint in the planning model.

Lu: I want to jump in here, Tom. What excites me is that they're using PDDL two point one, which is the standard language for automated planning. That means they're not building a one-off solver. They're encoding railway operations in a way that any temporal planner can read and solve. That's a huge deal for reproducibility.

Meng: But let me ask the practical question. How does this actually run in real time? Because a train disruption doesn't wait for a planner to think for thirty minutes.

Jane: That's a fair push, Meng. The paper reports that for smaller networks, the planners solve in under a second. For the very largest cases with a thousand track points and a hundred twenty trains, it can take up to thirty minutes. So there's definitely a gap between what's feasible offline and what's needed for real-time response.

Tom: Still, the fact that they can generate a complete, conflict-free operational plan with exact timestamps for every action — including turnout switches — is a big step beyond just producing a timetable. We'll dig into how they model all that next.

Summary: Tom: So we're back with "A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems." Jane, walk us through what the paper actually does, because I know our listeners want the meat of it.

Jane: The core idea is to treat the entire railway operation as a temporal planning problem. They define a set of durative actions — things like driving a train, switching a turnout, boarding passengers, attaching an auxiliary engine — and each action has a duration and conditions that must hold while it executes.

Lu: And the clever part is how they handle the safety constraints. When a train occupies a track segment, that segment becomes inaccessible in both directions. That's their mutual exclusion mechanism. It prevents two trains from ever colliding, and it's enforced by the planner automatically, not by a human dispatcher.

Tom: That's a big deal. Because in many real railway systems, the human operator is the one making those calls, and that's where errors creep in. The paper actually cites that human error accounted for over half of operational incidents in Bangladesh Railway in two thousand twenty-two-two thousand twenty-three.

Jane: Right. And they model four types of disruptions explicitly: blocked tracks, blocked trains, speed slowdowns, and engine failures. Each one has a recovery strategy encoded in the domain. For example, if a train's engine fails, an auxiliary engine gets dispatched, attaches to the train, and then moves it at the auxiliary engine's speed.

Meng: I'm curious about the benchmark they built. They mention two hundred instances. What does that actually look like?

Jane: They scaled it systematically. Four size categories — small, medium, large, very large — with up to a thousand track points and a hundred twenty trains. Half the instances are nominal, meaning no disruptions. The other half inject disruptions of increasing severity. That lets them test both scalability and robustness in a controlled way.

Tom: And the results? What did they find?

Jane: Both planners they tested — POPF and OPTIC — solved all hundred nominal instances and ninety-nine of the hundred disrupted ones. The one failure was the absolute largest case with a hundred eight concurrent disruptions. Every plan they generated passed validation with the VAL validator, which checks that all the temporal constraints actually hold.

Lu: What I find striking is that both planners produced identical makespans and identical delays. That suggests the problem is so tightly constrained by gauge compatibility and track exclusivity that there's essentially one optimal way to coordinate the trains. The solution space is narrow, and the planners find the same answer.

Meng: So the framework is sound, but the real question is whether thirty minutes of computation is acceptable when a train is stuck on a track right now. That's a deployment challenge, not just a research result.

Tom: Good point, Meng. And that's actually where the paper's future work section gets interesting — they talk about plan repair under uncertainty. We'll get into what they propose next.

Improvements: Tom: We're still on "A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems." Jane, we've covered what the paper does. Now let's talk about what it improves compared to what came before.

Jane: The biggest improvement is that they're not just generating a timetable. They're generating an executable action sequence. That means the output isn't just "train T1 arrives at station F at time eighty-seven." It's a full list of timestamped commands: drive this train on this segment, switch this turnout, board these passengers, attach this engine.

Lu: That's the shift from high-level planning to low-level control. Most prior work in railway scheduling uses mathematical optimization like MILP or MaxSAT to produce a schedule, but then a human dispatcher has to figure out how to actually execute it. This paper removes that gap by having the planner directly produce the operational commands.

Meng: So it's like the difference between a GPS telling you "arrive at the airport by five PM" versus a GPS telling you "turn left in three hundred meters, then merge onto the highway." One is a goal, the other is a set of instructions.

Jane: Exactly. And that's important for safety, because the instructions include the turnout switching. In their example network, the plan explicitly says when to switch junction B from connecting to D to connecting to C. That's a physical action that has to happen at a specific time, and the planner schedules it.

Tom: What about the disruption handling? How is that an improvement over existing methods?

Jane: They integrated disruption recovery directly into the planning domain. So when a track is blocked, the planner doesn't just wait — it can reroute trains through alternative gauge-compatible paths. When an engine fails, it dispatches a helper engine. These aren't separate modules bolted on afterward. They're part of the same planning problem, so the planner can optimize the whole recovery holistically.

Lu: And the delay decomposition they report is really informative. They break total delay into three components: slowdown delay, blockage delay, and engine failure delay. Slowdowns account for about sixty percent of the total delay, which makes sense because slowdowns affect every train on the affected segment, not just one.

Meng: So the improvement isn't just in the planning quality, but in the diagnostic power. You can look at the delay breakdown and see exactly which type of disruption is costing the most time. That's useful for infrastructure investment decisions — if slowdowns are the biggest driver, maybe the fix is track maintenance, not more trains.

Tom: That's a really practical take, Meng. And it connects to what the paper says about reducing dependence on manual decision making. The framework gives operators a clear picture of what's happening and what the plan is, rather than leaving everything to human judgment under pressure.

Jane: Right. And the correlation analysis they did shows that disruption severity scales predictably with delay. That means operators can estimate how bad a situation will get based on how many disruptions are active, which helps with resource allocation.

Conclusion: Tom: We've reached the end of our discussion on "A Temporal Planning Framework for Disruption Aware Dynamic Route Optimization in Heterogeneous Railway Systems." Jane, give us the final wrap-up.

Jane: This paper shows that temporal planning can handle the messy reality of railway operations — multiple gauges, different train speeds, disruptions of various types — and produce a complete, timestamped, conflict-free plan. It's a proof that automated planning isn't just for toy problems.

Lu: The benchmark design is a real contribution. Two hundred instances with controlled scaling, validated plans, and a clear delay decomposition. That gives the research community a solid foundation to build on and compare against.

Meng: And while the thirty-minute solve time on the largest cases is a limitation for real-time use, the framework's structure means it could be adapted for incremental re-planning. The future work on plan repair is the right direction.

Tom: So what does this mean for the world? If this kind of framework matures, railway operators could respond to disruptions faster and more safely, with less reliance on human dispatchers making split-second decisions.

Jane: And that matters because railways are one of the most efficient ways to move people and goods at scale. Making them more resilient to disruptions has a direct impact on daily life — commuters, freight, supply chains. It's not glamorous, but it's infrastructure that keeps society moving.

Tom: Well said. We'll be watching to see if they extend this to plan repair under uncertainty and real-time data integration. Thanks for joining us, and we'll see you on the next paper.

Jane: Goodbye, everyone.

Department of Computer Science and Engineering, Dhaka University of Engineering & Technology

cs.AI

Submitted: 2026-06-12

Updated: 2026-08-14

Code: https://github.com/PollobRay/Railway-Route-Planning

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: "existing studies predominantly focuses on high-level timetabling, omitting operational details such as track switching coordination.

Key concepts

Temporal Planning: A method of scheduling that determines not only what actions must be taken but also the exact timing for each action. It is more detailed than a simple to-do list, providing a complete schedule with timestamps for every task.
Heterogeneous Railway Systems: Railway networks where components are not uniform. This includes variations like different track gauges (the distance between rails), requiring specialized planning that accounts for physical compatibility constraints.
Disruption Aware Optimization: The ability of the planning system to account for unexpected events, such as blocked tracks, engine failures, or speed slowdowns. The framework integrates recovery strategies directly into the optimization process.
Executable Action Sequence: The output of the planner is a full list of timestamped commands (e.g., 'switch turnout B from D to C'). This moves beyond simple timetables by providing low-level instructions for actual operational control.

Terminology

Benchmark — 26× in this paper · explained in 2 episodes
Very Large — 7× in this paper
Planning Domain Definition Language — 4× in this paper
Robustness — 4× in this paper · explained in 5 episodes
Core — 3× in this paper · explained in 2 episodes
Mixed Integer Linear Programming — 2× in this paper
Resource Constraints — 2× in this paper · explained in 3 episodes
Abstraction — 1× in this paper · explained in 1 episode
Alignment — 1× in this paper · explained in 1 episode
Computational Tractability — 1× in this paper · explained in 2 episodes
Disruption Aware Railway Temporal Planning — 1× in this paper
Pruning — 1× in this paper · explained in 4 episodes
Temporal Relaxed Planning Graph — 1× in this paper

Summary

Summary

The paper proposes a temporal planning-based framework for dynamic route optimization and disruption management in heterogeneous railway systems, addressing gaps in existing research that predominantly focuses on high-level timetabling while omitting operational details such as track switching coordination. The authors state: "existing studies predominantly focuses on high-level timetabling, omitting operational details such as track switching coordination. As a result leaving decision to human operators, increasing safety risks into railway operations."

The framework formulates railway operations as a temporal planning problem using PDDL 2.1, explicitly modeling gauge compatibility constraints and diverse disruption scenarios. The authors note that multi-gauge networks amplified this complexity by introducing multiple gauge types for track and train and require matching gauge types for train movement, citing examples such as Spain (Iberian gauge 1668 mm, standard gauge 1435 mm, dual-gauge) and Bangladesh (meter gauge 1000 mm, broad gauge 1676 mm, dual gauge). The framework generates conflict-free timestamped operational plans specifying both optimized schedules and executable action sequences.

The paper contributes three main elements: (1) a temporal planning domain for heterogeneous railways encoding gauge compatibility and varying train profiles, (2) integrated disruption management handling four disruption types—blocked tracks, blocked trains, engine failures, and speed slowdowns—with formal recovery strategies, and (3) a benchmark problem set of 200 instances (100 nominal and 100 disrupted) scaling from small networks up to 1,000 track points and 120 trains.

The domain model defines five object types (track-point, train, gauge-type, engine, station) and ten durative actions categorized into movement operations (drive-train, drive-engine), passenger service operations (board-passengers), disruption recovery operations (drive-engine-to-damaged-up-train, drive-engine-to-damaged-down-train, attach-engine, drive-assisted-train, resolve-train-blockage, clear-blocked-track), and infrastructure operations (turnout). The drive-train action duration is computed as distance divided by train speed adjusted for slowdown factors, and includes temporal conditions ensuring gauge compatibility, track clearance, and destination availability throughout execution.

The disruption taxonomy formalizes: (1) blocked track disruption requiring waiting or rerouting via gauge-compatible alternative paths using turnout operations, (2) blocked train disruption where the affected train remains immobilized while other trains continue, (3) slowdown disruption where travel time is computed as track-distance divided by (1-α) times train-speed with α denoting the slowdown factor, and (4) engine failure disruption requiring dispatch of auxiliary engines and coupling operations.

The benchmark problem set is organized into four scale categories (Small, Medium, Large, Very Large) with three progressively harder sub-levels each. Nominal-operation instances have zero disruptions, while disrupted instances combine infrastructure scaling with increasing disruption severity—blocked trains, blocked tracks, engine failures, slowdown segments, and limited auxiliary engines. The authors emphasize the scaling is controlled and monotonic with disruption severity increasing proportionally with network size.

Experimental evaluation employed two state-of-the-art temporal planners, POPF and OPTIC, with a 30-minute time limit per instance, running on a workstation with a 7th-generation Intel Core i7 processor and 32 GB RAM. All generated plans were validated using the VAL plan validator, which confirmed every plan satisfies all temporal constraints, precondition requirements, and goal conditions specified in the domain with no mutex violations or goal achievement errors detected.

Results show both planners produced identical makespan, plan length, and total delay values across all solved instances, which the authors attribute to OPTIC extending POPF with the same TRPG heuristic, and to the tightly constrained solution space imposed by gauge compatibility and single-track mutual exclusion. For nominal instances, makespan remained stable (mean 34.15±1.85) with zero total delay, while plan length grew proportionally with network size (from 15.00±6.10 for S1 to 226.60±8.60 for VL3). For disrupted instances, all metrics increased monotonically: makespan from 42.52±19.44 (S1) to 57.21±21.45 (VL3), plan length from 18.00±6.10 to 251.10±7.70, and total delay from 28.00±42.16 to 377.41±580.97.

Both planners solved all 100 nominal instances and 99 of 100 disrupted instances, with the single failure being the largest configuration (120 trains, 1,000 track points, 108 concurrent disruptions). Computation time ranged from approximately 0.06 seconds for small nominal instances to around 45 seconds for very large nominal instances, and from 0.28 seconds to 619.00 seconds for disrupted instances, with the VL3 group plateauing due to the 30-minute time limit.

The paper introduces three domain-specific delay metrics: Total Delay (cumulative deviation between actual and ideal arrival times over all trains and track points), Slowdown Delay (extra time from reduced-speed segments), Blockage Delay (cumulative duration of resolve-train-blockage and clear-blocked-track actions), and Engine Failure Delay (sum of dispatch and attach durations for auxiliary engines). Delay decomposition reveals slowdowns contribute 60.1% of total delay, engine failures 30.4%, and blockages 9.5%, with slowdown share growing in larger networks where higher disruption density leaves less room to reroute around affected segments.

Statistical analysis using Spearman rank correlation confirmed perfect positive correlations (ρ = 1.000, p < 0.001) between all system size metrics and computation time, validating controlled complexity scaling. Disruption characteristics exhibited very strong to perfect positive correlations with total delay (ρ ≥ 0.976, p < 0.001), demonstrating that coordination degradation scales proportionally with disruption intensity across the benchmark.

The authors conclude that temporal planning as a practical and scalable foundation for intelligent railway management systems and outline future work including plan repair under uncertainty, enabling adaptive re-planning in response to incomplete or evolving operational information, as well as derailment modeling and integration with real-time data streams. The planning domain, problem instances, results, and validation outputs are available at https://github.com/PollobRay/Railway-Route-Planning.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

Improvement: Encode railway operations as PDDL 2.1 durative actions with explicit gauge compatibility constraints, turnout operations, and passenger boarding activities.
What it can do: Generate timestamped, conflict-free operational plans specifying exact action sequences (drive-train, turnout, board-passengers) with precise start times and durations, rather than just high-level timetables.
Improvement: Model four disruption types (blocked tracks, blocked trains, slowdowns, engine failures) as first-class predicates with dedicated recovery durative actions (clear-blocked-track, resolve-train-blockage, attach-engine, drive-assisted-train).
What it can do: Automatically generate recovery plans that reroute trains via gauge-compatible alternatives, dispatch auxiliary engines, and compute slowdown-adjusted travel times using the formula distance / (speed × (1 − α)).
Improvement: Model junction switching as a durative action with explicit turnout-alternatives predicates and switching time, eliminating reliance on human operators.
What it can do: Automatically decide when and how to switch track connections at junctions to enable rerouting, while ensuring no train occupies the affected segment during switching.
Improvement: Encode train-gauge, track-gauge, and engine-gauge predicates to restrict train movements to physically compatible tracks (e.g., meter-gauge trains only on meter or dual-gauge tracks).
What it can do: Generate feasible routes that respect infrastructure constraints, preventing invalid movements that would be impossible in real-world multi-gauge networks (e.g., Spain, Bangladesh).
Improvement: Encode free, track-accessible, and over all conditions to prevent simultaneous occupancy of track segments and points.
What it can do: Guarantee collision-free schedules where no two trains occupy the same track point at the same time, verified by the VAL validator across all 199 solved instances.
Improvement: Use the 200-instance benchmark with monotonic scaling (50–1,000 track points, 3–120 trains) and Spearman correlation analysis to predict planner performance.
What it can do: Predict computation time and delay growth with high confidence (ρ = 1.000 for size vs. time; ρ ≥ 0.976 for disruptions vs. delay), enabling resource allocation decisions for real-time deployment.
Improvement: Decompose total delay into slowdown, blockage, and engine-failure components using domain-specific definitions.
What it can do: Identify that slowdowns contribute 60.1% of delays, engine failures 30.4%, and blockages only 9.5%, enabling targeted infrastructure investment and disruption mitigation strategies.
Improvement: Design a tightly constrained domain where both POPF and OPTIC converge to identical solutions (makespan, plan length, delay).
What it can do: Provide confidence that the generated plans are near-optimal and reproducible, regardless of which temporal planner is deployed in production.

Generate executable railway operation plans — not just timetables, but exact action sequences with timestamps, ready for direct deployment to signaling systems.
Automatically resolve disruptions in real time — when a track is blocked, the system reroutes trains via alternative gauge-compatible paths; when an engine fails, it dispatches and attaches auxiliary engines.
Handle multi-gauge networks safely — the system never proposes a route that violates gauge compatibility, preventing physical impossibility errors.
Coordinate concurrent train movements — the system schedules overlapping actions (e.g., train T1 moving while T2 boards passengers) while maintaining safety invariants.
Predict operational delays before execution — using the delay decomposition, the system can forecast which disruption types will cause the most impact and prioritize mitigation.
Scale from small to very large networks — from 50 to 1,000 track points and up to 120 trains, with predictable computation time growth (0.06s to 10 minutes).
Validate plans automatically — every generated plan is verified by VAL, ensuring temporal consistency, precondition satisfaction, and goal achievement before deployment.
Reduce human error in railway operations — by automating turnout decisions and disruption recovery, the system addresses the 50.77% of incidents caused by human error in Bangladesh Railway (2022–2023).

Abstract

Efficient route optimization play a vital role in ensuring both safety and punctuality in railway operations. It is very crucial particularly in heterogeneous multi-gauge railway networks with varying train speed, stopping pattern, infrastructure compatibility constraints increase coordination complexity. In single-track systems these challenges are further intensify due to all trains to share the same track and requires frequent track switching.Stochastic disruptions events including blocked tracks, blocked trains, engine failure and speed slowdowns introduces additional unpredictability in operations and deviate the timetable. However, existing studies predominantly focuses on high-level timetabling, omitting operational details such as track switching coordination. As a result leaving decision to human operators, increasing safety risks into railway operations. This study proposes a framework based on temporal planning for dynamic route optimization and disruption management in heterogeneous railway systems. The framework formulates railway operations as a temporal planning problem using PDDL 2.1 with explicitly modeling gauge compatibility constraints and diverse disruption scenarios. It generates conflict-free timestamped operational plans specifying both optimized schedules and executable action sequences. To evaluate the proposed framework, we developed a benchmark problem set with 200 instances using up to 1,000 track points and 120 trains. Two state-of-the-art temporal planners and a plan validator were employed to assessed the framework. The experimental results demonstrate that the framework effectively generates temporal operational plans for heterogeneous railway systems and handles multi-gauge constraints, disruptions, and reduces dependence on manual decision making.