From Fast Routes to Safe Shelters: Rethinking Humanitarian Supply-Chain Resilience

When disaster strikes, humanitarian planners rarely face a simple question of how to move aid faster. They must decide who should receive help first, which damaged routes remain usable, how much delay can be tolerated and whether already strained relief centres can absorb more people. Every choice protects some priorities while weakening others.

A new study, Resilient Last-Mile Allocation in Humanitarian Supply Chains: A Reproducible Four-Objective Fuzzy Framework, published in Applied Sciences, attempts to make those competing pressures visible rather than burying them inside a single efficiency score. Authored by Peyman Rabiei, Daniel Arias-Aranda and Vladimir Stantchev, the research develops an open-source decision-support model that evaluates humanitarian allocation through four dimensions of resilience: Resourcefulness, Robustness, Rapidity and Redundancy.

The framework combines fuzzy reasoning, rules expressed in terms such as high vulnerability, dangerous routes or rapidly depleting resources, with multi-objective optimisation. Instead of producing one supposedly ideal answer, it generates a range of defensible options showing what planners gain and sacrifice under each allocation.

The First Hours Turn Logistics Into Triage

Humanitarian logistics is often described as a problem of delivery, but the earliest phase of a disaster is closer to triage. Available transport, shelter capacity and supplies are usually smaller than immediate need. Planners must therefore decide whom to direct to safety, where to send them and which risks to accept.

The study begins from this scarcity. Its scenarios require the model to choose only a subset of affected people for immediate allocation. A person's vulnerability, local infrastructure damage and remaining access to essential resources influence their priority. Road conditions, hazards and travel time shape the feasibility of reaching a relief centre. Occupancy and resource depletion determine whether a centre can continue receiving people without becoming the next point of failure.

Conventional optimisation models frequently reduce such decisions to cost, distance, total travel time or the number of people left without assistance. Those indicators are valuable, but they can obscure the logic used by civil-protection agencies, which often think in terms of system resilience.

The new framework maps each objective onto the established 4R model. Resourcefulness asks whether the most vulnerable people are being identified and prioritised. Robustness considers whether routes can function under damaged or hazardous conditions. Rapidity measures whether people can arrive in time. Redundancy assesses whether demand is distributed across relief centres instead of concentrated in facilities already close to exhaustion.

This vocabulary matters because emergency planners should not have to translate an abstract optimisation output into the language of operational policy while a crisis is unfolding. A model framed around resilience enables them to ask direct questions: Is the fastest option too fragile? Is the safest route too slow? Are vulnerable people being redirected to a centre that may soon run out of capacity?

The model does not combine these concerns into one weighted answer. It generates a Pareto front—a set of options in which improving one objective would worsen another. This is closer to the reality of disaster response, where a safer decision is not always faster, and the fairest allocation may not be the easiest to execute.

Splitting Speed from Safety Reveals the Real Choice

The study's most consequential modelling decision is to separate transport safety from transport speed. Earlier versions of the framework combined road condition, hazard exposure and travel duration into one transportation-infeasibility score. The approach was computationally convenient, but it fused two fundamentally different questions: whether a route is likely to remain usable and whether it can deliver someone quickly enough.

The four-objective model treats them as independent dimensions. The researchers tested this distinction across 270 four-objective runs and found an average Spearman rank correlation of –0.29 between Robustness and Rapidity. In 82% of the generated solution fronts, the two measures met the study's threshold for carrying non-redundant information.

The negative relationship is revealing. Safer routes often required more time, while faster routes sometimes exposed people to greater danger. This is not a technical inconvenience; it is the operational dilemma itself.

A planner may reasonably select a hazardous but rapid route when a critically injured person must reach care within minutes. In another case, the same planner may choose a longer but more reliable route because the danger of infrastructure failure outweighs the delay. A combined score would conceal how that compromise was made. Separate indicators allow the decision-maker to choose deliberately.

The study found that roughly half of the four-objective solutions became dominated when compressed into a three-objective format. However, the authors tested this result against random combinations of speed and safety and found that most of the apparent loss arose simply from reducing four dimensions to three.

The qualification strengthens rather than weakens the paper. The authors do not claim that the new formulation magically creates superior solutions. Its advantage is visibility. Any single transport score must embed one predetermined balance between speed and safety. The four-objective framework leaves that balance open for human judgment.

Faster Algorithms Help, but Better Data Matters More

The researchers compared three evolutionary algorithms, NSGA-II, NRGA and NSGA-III, across small, medium and large synthetic scenarios. The largest involved 300 affected people, ten relief centres and the immediate allocation of 100 people. In total, the study conducted 540 experimental runs.

NSGA-III emerged as the fastest algorithm. It reduced computation time relative to NSGA-II by 37% in the small scenario, 32% in the medium scenario and 27% in the large scenario, producing an average speed advantage of 32%.

The result has obvious appeal in emergency management, where time-to-decision matters. Yet the solution-quality comparison was less straightforward. Under one fixed reference point, NSGA-II achieved higher hypervolume, a measure used to assess the breadth and quality of multi-objective solutions. When the researchers applied an instance-specific reference point, NSGA-III moved ahead, but the difference was not statistically significant. NSGA-III generated broadly comparable decision sets more quickly. The study does not establish that it always produces better allocations.

The sensitivity analysis adds another layer of confidence. When the authors changed expert-assigned weights by plus or minus 20%, variation in the four mean objectives remained below 1% across all tested problem sizes. This suggests that modest disagreements among experts about the relative importance of age, disability, injury or road hazards may not dramatically change the resulting recommendations.

Most of the fuzzy rule systems were also stable when individual rules were removed. The exception was the three-rule Rapidity classifier, which categorised travel time only as short, moderate or long. Removing one rule eliminated one-third of that entire rule base. Median changes in solution quality exceeded the study's preferred threshold in the small and medium scenarios, while the most disruptive deletion produced a change approaching 25%.

The authors propose expanding the classifier to five levels. That is more than a technical adjustment. It shows why transparent rule-based systems are valuable: weaknesses can be identified, explained and revised.

Still, algorithmic speed and model stability should not distract from a larger constraint. The framework can only be as reliable as the information entering it. During real disasters, road conditions change rapidly, mobile networks fail, shelter occupancy figures become outdated and vulnerability records may be incomplete. A sophisticated model operating on delayed or biased information can produce precise but misleading recommendations.

A Strong Framework Still Awaits Its Real-World Test

The paper links resilience theory to an operational allocation model, separates ethically meaningful trade-offs, tests multiple algorithms, examines sensitivity and releases an openly licensed Python implementation. Reproducibility is treated not as an academic formality but as part of resilience itself: agencies and researchers can inspect the logic, reproduce the experiments and adapt the tool without depending entirely on a proprietary vendor.

This could be particularly valuable in developing countries, where disaster exposure is high and access to expensive commercial systems may be limited. A tool that runs on ordinary hardware and uses open code could support local adaptation, regional training and shared emergency-planning infrastructure.

Its possible applications also connect to broader development goals. Better last-mile allocation could support health protection, disaster resilience, inclusive infrastructure and institutional preparedness. The framework may also prove useful beyond sudden-onset disasters, including heatwaves, displacement emergencies, epidemics and climate-related evacuations.

However, the study remains a computational proof of concept. All three problem sizes use synthetic data rather than information from an actual disaster. The model has not yet been tested against observed emergency decisions, real travel networks or documented outcomes. The fuzzy rules were not produced through a formal multi-expert consensus process involving a diverse group of humanitarian practitioners. The model also focuses primarily on road transport, while many real emergencies depend on walking routes, boats, public transport, motorcycles or informal community networks. Shelter capacity is represented through penalties rather than always enforced as a hard operational constraint.

Future research should test the framework against documented events, compare its recommendations with manual planning and simpler allocation methods, and examine whether it improves outcomes under changing field conditions. The authors identify the 2023 Türkiye–Syria earthquake and the 2024 Valencia floods as possible case studies. Formal consultations with civil-protection officials, humanitarian organisations and affected communities would also strengthen the legitimacy of the rules.