💻 computer science

Risk-Aware Planning for Transit Desert Remediation Under Demand Uncertainty

This paper proposes a risk-aware planning framework using a partially observable Markov decision process with Conditional Value-at-Risk constraints to remediate transit deserts under demand uncertainty, demonstrating through a multi-city evaluation that this adaptive approach significantly outperforms static optimization in restoring transit access while managing financial tail risk.

Original authors: Polina Khoroshevskaya, Ashish Kumar Perukari

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Polina Khoroshevskaya, Ashish Kumar Perukari

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a city planner trying to fix "transit deserts"—neighborhoods where people desperately need a bus or train, but none exists. The big problem is a catch-22: You don't know how many people will actually ride the bus until you build it. If you build a bus line and no one shows up, you've wasted millions of dollars. If you don't build it, people remain stranded.

This paper proposes a smart, step-by-step strategy to solve this guessing game. Instead of trying to predict the future perfectly (which is impossible), the authors treat the problem like a game of exploration with a safety net.

Here is the breakdown of their approach using simple analogies:

1. The Problem: Planning in the Dark

Usually, cities try to guess where to put new bus lines using surveys and old data. But in a "transit desert," there is no history to look at. It's like trying to guess how many people will buy ice cream in a town that has never had an ice cream shop.

The Risk: If you guess wrong and build a line nobody uses, you lose money (financial risk).
The Goal: You want to help the most people (especially low-income residents) without going broke.

2. The Solution: The "Smart Explorer" (POMDP)

The authors created a computer model called a Partially Observable Markov Decision Process (POMDP). Think of this as a smart explorer with a map that updates itself.

The Starting Map (The Prior): Before building anything, the computer looks at clues like population density, how many people own cars, and how far they live from existing buses. It makes an educated guess (a "prior") about how many people might ride.
The Safety Net (CVaR): The model has a "risk alarm." It is programmed to avoid scenarios where the city could lose a huge amount of money. It won't gamble on a high-risk neighborhood just because the potential reward looks high; it keeps the city's budget stable.
The Learning Loop: This is the magic part.
1. The planner picks a few neighborhoods to build a bus line in first.
2. Once the bus runs, the city sees the actual number of riders.
3. The computer updates its map: "Okay, we guessed 100 riders, but 200 showed up! Let's adjust our guess for the next neighborhood."
4. With this new, better information, the planner decides where to build the next bus line.

3. The Strategy: "Myopic" but Smart

The authors tested a strategy called a "myopic belief-aware planner."

"Myopic" sounds bad, but here it just means the planner focuses on making the best decision right now rather than trying to solve the whole 5-year puzzle at once.
"Belief-aware" means it remembers that its current guesses might be wrong. It values information. It might choose to build a bus in a neighborhood where it's unsure of the demand, just to find out the truth, so it can make better choices later.

4. The Results: Testing in 25 Cities

The team tested this "Smart Explorer" against two other methods in 25 cities across the Americas (including Chicago, Detroit, Austin, and Bogota):

Static Planning: The old way. Guess once, build everything based on that guess, and never change the plan.
Greedy Planning: Just build the cheapest lines first.
Random: Just picking spots at random (the control group).

The Findings:

Better Coverage: The "Smart Explorer" fixed 53.6% of the transit deserts on average, which was 5% better than the old "Static" method. That might sound small, but in city planning, that's a huge difference in how many people get access to jobs and schools.
When it Works Best: The strategy shined when the city had a moderate budget. If the budget was tiny, there wasn't enough money to learn anything. If the budget was huge, they could just build everywhere anyway. The "learning" strategy was most valuable when they had to be careful with their money.
Robustness: Even when the initial guesses were way off (up to 50% wrong), the Smart Explorer still did better than the Static planner because it could correct its course as it got real data.
The Cost Factor: They found that population density is the biggest predictor of how much it costs to fix a transit desert. Dense cities are cheaper to fix because you can run one bus route and pick up many people. Sprawling, low-density cities are expensive because you have to drive long routes to find a few riders.

5. The Bottom Line

The paper argues that cities shouldn't treat their 5-year bus plan as a rigid, unchangeable document. Instead, they should treat it like a learning process.

By building a few lines first, watching who rides them, and then using that real-world data to decide where to build next, cities can fix more transit deserts, help more low-income residents, and avoid wasting money on empty bus routes. The "Smart Explorer" doesn't need to be a psychic; it just needs to be willing to learn from its mistakes and successes as it goes.

Technical Summary: Risk-Aware Planning for Transit Desert Remediation Under Demand Uncertainty

1. Problem Definition

The paper addresses the challenge of transit desert remediation: allocating scarce capital to improve public transportation in areas where demand exists but service is inadequate. The core difficulty lies in demand uncertainty. Unlike established corridors where ridership is observable, transit deserts lack historical data; agencies cannot observe ridership before service exists.

Traditional planning relies on four-step travel demand models using revealed preferences (surveys) and land-use forecasts. However, these models fail in unserved zones because they extrapolate from existing infrastructure, often reinforcing gaps by directing investment only to areas where demand is already certain. The authors frame this as a sequential decision-making problem under partial observability, where an agency must balance immediate equity gains against the risk of financial loss and the value of learning latent demand through initial service deployments.

2. Methodology

A. Formalization as a POMDP

The problem is modeled as a Partially Observable Markov Decision Process (POMDP) over a five-year horizon:

State ( $s_t$ ): Includes the current coverage map ( $v_t$ ), remaining budget ( $B_t$ ), and a hidden latent demand vector ( $d$ ). Demand is drawn once from a prior and remains fixed (epistemic uncertainty), rather than fluctuating as a stochastic shock.
Action ( $a_t$ ): Allocation of budget across census-tract zones. Coverage in a zone changes only when cumulative investment exceeds a specific threshold cost.
Observation ( $o_t$ ): Noisy ridership counts are observed only in zones that have received service. Unserved zones provide no signal, forcing the planner to "buy" information through deployment.
Reward: A composite function combining:
1. Equity-weighted coverage gain: Weighted inversely by median income ( $w_z \propto 1/\text{Income}_z$ ).
2. Risk Penalty: A Conditional Value-at-Risk (CVaR) term penalizing the expected financial shortfall in the worst-case $\alpha$ -fraction (set to 10%) of demand scenarios. This prevents the planner from chasing high-variance, high-reward zones at the expense of fiscal stability.

B. Belief Dynamics and Side Information

The system maintains a log-normal prior over latent demand for each zone, conditioned on side information (population density, employment, income, vehicle access, and distance to existing transit).

Prior Variance: Zones farther from existing transit are assigned higher prior variance, reflecting weaker evidence about ridership.
Bayesian Update: As service is deployed and ridership is observed, the posterior distribution is updated using a normal-normal conjugate update in log-space. Variance contracts with each observation, allowing the planner to refine demand estimates for served zones while remaining uncertain about unserved ones.

C. Solution Approach

Given the five-year horizon and moderate action space, the authors employ a myopic belief-aware planner rather than a full POMCP (Partially Observable Monte Carlo Planning) approach.

Objective: Maximize expected equity-weighted coverage gain minus the CVaR penalty, plus an explicit Information Value (IV) bonus. The IV bonus incentivizes investing in zones with high posterior uncertainty to accelerate learning.
Optimization: The problem is solved via Sample-Average Approximation (SAA). The planner draws 200 demand scenarios from the current belief state and solves a Linear Program (LP) with auxiliary CVaR variables. This allows for fast computation (under 0.5 seconds per period per city).

D. Financial Model

The model integrates a unified financial framework for 25 cities, calculating:

Operating Cost: Based on vehicle-hours and regional price parity.
Capital Cost: Amortized costs for vehicles (12 years) and infrastructure (30 years).
Revenue: Estimated as ridership $\times$ fare $\times$ capture rate.
Net Subsidy: The difference between costs and revenue, used to calculate the CVaR risk metric.

3. Experimental Setup

Dataset: 25 cities across North and South America (including Chicago, Austin, Detroit, Portland, Bogotá, and others), varying in size, urban form, and transit maturity.
Zones: Census tracts (or administrative units in Bogotá), with 60–200 zones per city.
Baselines:
1. Static Optimization: A deterministic five-year plan allocating the entire budget based on the prior mean demand, without sequential updating.
2. Greedy Coverage: Funds the uncovered zone with the highest equity-to-cost ratio, ignoring risk and information value.
3. Random Allocation: Uniform random selection.
Metrics: Fraction of transit-desert tracts remediated ( $\Delta C$ ), cost efficiency, equity score, and financial risk (CVaR of net subsidy).

4. Key Results

A. Performance Gains

Remediation Rate: The POMDP planner remediates a median of 53.6% of transit-desert tracts over five years.
Improvement over Static: On average, the POMDP approach improves remediation rates by 5.0 percentage points (an 11.3% relative gain) compared to static optimization. It outperforms static planning in 16 of 25 cities.
Budget Sensitivity: Gains are most pronounced at moderate budgets. At 50% of the baseline budget, the advantage peaks at +9.9 percentage points. At 200% budget (where both policies saturate near 93% coverage), the gap narrows to +2.6 points.
Robustness: The planner maintains its advantage even when the prior demand model is significantly miscalibrated (up to 50% noise), whereas static planning performance remains flat because it cannot revise its initial forecast.

B. Structural Predictors

Cost Drivers: Population density is the strongest predictor of per-tract remediation cost ( $R^2 = 0.41$ ). Doubling population density reduces per-tract costs by approximately 30%.
Cost Efficiency: Compact, high-density cities (e.g., Chicago, San Francisco) are cheaper to remediate via route extensions, while sprawling low-density metros (e.g., Houston, Phoenix) require longer routes with lower ridership density, increasing costs.

C. Risk and Equity

Financial Risk: While the aggregate net subsidy risk (CVaR) is slightly higher for the POMDP planner (due to serving more zones overall), the risk per remediated tract is lower. The CVaR penalty effectively steers the planner away from high-uncertainty, high-failure-risk investments.
Equity: The equity-weighted reward ensures the adaptive strategy does not drift away from low-income zones; in highly uncertain cities, it actively shifts investment toward them.

D. Ablation Studies

Information Value: Removing the explicit Information Value (IV) bonus ( $\eta=0$ ) reduces performance by only 0–4 percentage points in deep-dive cities. This indicates that the primary driver of success is Bayesian belief updating on observed ridership, rather than an explicit exploration reward. The planner naturally learns from early deployments, improving later allocations.

5. Significance and Claims

The paper claims to be the first work to:

Formulate transit desert remediation explicitly as a POMDP to handle the "cold start" problem of unobservable demand.
Integrate CVaR constraints to manage financial tail risk in sequential public transport planning.
Provide an empirical benchmark across 25 diverse cities using a unified financial model, moving beyond illustrative examples.

The authors argue that their results demonstrate that early service deployments can be designed to gather information as well as to expand coverage. A five-year capital plan need not be fixed entirely on initial projections; instead, agencies can use early ridership observations to inform subsequent allocations without abandoning equity goals. The study suggests that belief-aware planning is most valuable in cities with uncertain demand and moderate budgets, offering a robust alternative to static, deterministic planning.

6. Limitations

The authors acknowledge several modeling simplifications:

The demand model is log-linear and does not capture induced demand (new riders generated by new service).
Cost estimates rely on national averages where city-specific data is unavailable.
Realized demand is treated as static across the planning horizon.
The action space omits political and regulatory constraints that often dictate actual service expansion.