Risk-Averse Stochastic User Equilibrium on Uncertain Transportation Networks

This paper proposes a convex, risk- and ambiguity-aware Stochastic User Equilibrium framework that integrates tail-risk management and distributional robustness to optimize traffic assignment on transportation networks disrupted by extreme weather events, demonstrating how these approaches fine-tune flow redistribution without subverting equilibrium choices.

The Gist

Imagine you are planning a daily commute to work. Usually, you pick the route that gets you there the fastest on average. But what happens when a massive storm hits? Suddenly, the "fastest" road might be flooded, turning a 20-minute drive into a 2-hour nightmare.

This paper tackles a very real problem: How do we plan traffic routes when the weather is unpredictable and disasters can happen?

The authors argue that traditional traffic models are like a person who only looks at the average weather. They say, "On average, it's sunny, so I'll take the highway." But a risk-averse person knows that even if it's sunny 90% of the time, that 10% chance of a flood is terrifying. They would rather take a slightly longer, safer backroad than risk getting stuck in a flood.

Here is the breakdown of their solution, using simple analogies:

1. The Problem: The "Average" Trap

Traditional traffic models calculate the "expected" travel time.

• The Analogy: Imagine a coin flip. If heads, you get there in 10 minutes. If tails, you get stuck for 10 hours. The average time is 5 hours and 5 minutes. A traditional model says, "That's fine, let's go!"
• The Reality: A risk-averse driver looks at that 10-hour tail and says, "No way! I'd rather take a 45-minute route that is always safe, even if it's usually slower." The paper shows that ignoring these "worst-case" scenarios leads to bad decisions when disasters strike.

2. The Solution: A "Safety-First" Calculator

The authors created a new mathematical framework called Risk-Averse Stochastic User Equilibrium (TSUE). Think of this as a new GPS algorithm that doesn't just look at the average time, but also worries about the "worst-case" scenario.

They use two main ingredients to build this calculator:

• The Mean (The Average): How long the trip usually takes.
• The CVaR (The "Bad Day" Penalty): This stands for Conditional Value at Risk. It's a fancy way of asking, "If things go wrong, how bad will it be on average?"

They combine these into a single "Safety Score." You can adjust a dial (called $\lambda$) to decide how much you care about the bad days.

• Dial set to Low: You are like a casual driver; you mostly care about the average speed.
• Dial set to High: You are a paranoid driver; you will take a much longer route just to avoid even a tiny chance of a flood.

3. The Twist: What if we don't know the odds?

Here is the tricky part. In a real flood, we might not know the exact probability. Is there a 5% chance of flooding, or a 7%? Traditional models break if your guess about the probability is slightly wrong.

The authors added a layer of Distributionally Robust Optimization (DRO).

• The Analogy: Imagine you are betting on a horse race.
  • Standard Model: You bet based on the horse's past stats. If the stats are slightly off, you lose.
  • Robust Model: You assume the stats might be slightly wrong. You ask, "What is the worst possible version of these stats that could still be true?" and you bet based on that.
• The Result: This makes the traffic plan "bulletproof." Even if the weather forecast is slightly inaccurate, the traffic flow won't collapse. It finds a route that is safe even in the "worst-case" version of reality.

4. How it Works in Practice (The Chicago Test)

The authors tested this on a model of downtown Chicago.

• The Scenario: They simulated a flood that could close major roads.
• The Result:
  • Old Way (Average only): Traffic stayed on the main highways. When the flood hit, everyone got stuck.
  • New Way (Risk-Averse): The system automatically shifted traffic before the flood. It sent cars onto smaller, safer side streets (radial routes) that were slightly longer but wouldn't flood.
  • The "Robust" Way: This was the smartest. It didn't just shift traffic; it shifted it more aggressively to the safest routes, ensuring that even if the flood was worse than predicted, the network wouldn't gridlock.

5. Why This Matters

The paper proves that by tweaking how we calculate "cost" (adding a fear of the worst-case), we can:

1. Prevent Gridlock: By spreading cars out to safer, less obvious routes before a disaster hits.
2. Be More Resilient: The traffic flow doesn't change wildly if our weather data is slightly wrong.
3. Save Time: While the routes might be slightly longer on a sunny day, they save massive amounts of time during a crisis.

In a nutshell: This paper teaches us how to build a traffic system that doesn't just hope for the best, but prepares for the worst, keeping cities moving even when the sky turns gray and the roads start to flood. It's about trading a little bit of speed for a lot of peace of mind.

Technical Summary

1. Problem Statement

The paper addresses the challenge of traffic assignment in transportation networks disrupted by extreme weather events (e.g., flooding, heavy rain). These hazards introduce regime-dependent uncertainty, causing link capacities to drop, speeds to decrease, and roadways to close.

• Limitation of Current Models: Conventional Stochastic User Equilibrium (SUE) models typically rely on the expected value (mean) of travel time. This approach fails to account for "rare but severe" delays (long-tailed distributions) that characterize hazard scenarios. Risk-neutral assignment underestimates the disutility of extreme delays, leading to suboptimal route choices where travelers might select routes that are fast on average but prone to catastrophic failure.
• The Gap: There is a need for a framework that captures tail risk (reliability) and distributional ambiguity (uncertainty in the probability distribution itself due to limited data or regime shifts) while maintaining mathematical tractability (convexity) for equilibrium computation.

2. Methodology

The authors propose a Risk-Averse Truncated Stochastic User Equilibrium (TSUE) framework that integrates behavioral risk aversion with distributionally robust optimization.

A. Core Framework: Normalized Mean-CVaR Certainty Equivalent

Instead of using pure Expected Value or pure Conditional Value at Risk (CVaR), the authors introduce a normalized mean-CVaR certainty equivalent ($\Phi_{\alpha,\lambda}$) to represent travelers' perceived costs:
$$\Phi_{\alpha,\lambda}(Y) = (1 - \bar{\lambda})E[Y] + \bar{\lambda}CVaR_{\alpha}(Y)$$

• Parameters:
  • $\alpha$: The confidence level defining the tail threshold (e.g., 95th percentile).
  • $\lambda$: A behavioral parameter controlling the degree of risk aversion.
  • $\bar{\lambda}$: An effective mixing weight derived from $\alpha$ and $\lambda$ to ensure the model preserves convexity and scales correctly with deterministic travel times.
• Truncated Logit: The model employs a truncated logit choice mechanism. Paths with perceived costs exceeding a reliability threshold ($\pi$) are assigned zero probability, effectively filtering out unreliable routes.

B. Two Formulations (TSUE-SP)

The paper develops two complementary approaches to integrate this risk metric into the equilibrium:

1. Approach A (Path-Based): Applies the certainty equivalent directly to each path's travel time. While intuitive, this breaks link-additive separability, generally requiring a fixed-point iteration (Variational Inequality) rather than a convex optimization.
2. Approach B (Potential-Based): Applies the risk functional to the system-level potential (congestion + entropy).
   • B1: Risk applied to the total potential (congestion + entropy).
   • B2: Risk applied only to the congestion potential, keeping the entropy term outside.
   • Key Finding: Under specific conditions (e.g., regime-dependent ordering of travel times), Approach A and Approach B yield equivalent equilibrium flows. Approach B allows for a single-level convex reformulation using the Rockafellar-Uryasev (RU) representation of CVaR, making it computationally tractable.

C. Distributionally Robust Optimization (TSUE-DRO)

To address the uncertainty in the probability distribution itself (calibration errors, limited data, regime shifts), the authors extend the framework to Distributionally Robust Optimization (DRO).

• Ambiguity Set: They utilize a 1-Wasserstein ambiguity set ($\mathcal{D}$) centered around an empirical distribution. This allows the model to optimize against the worst-case distribution within a specific distance ($\rho$) from the nominal data.
• Regime-Dependent Sets: The model is extended to handle piecewise-stationary hazards (e.g., Normal Rain vs. Heavy Rain vs. Flooding). It distinguishes between:
  • Scenario A: Robustness against the mixture distribution (unconditional risk).
  • Scenario B: Regime-conditional robustness (informed travelers who know the current weather regime).
• Algorithm: The DRO problem is reformulated as a tractable Second-Order Cone Program (SOCP) or solved via a Benders decomposition algorithm with cutting planes for semi-infinite constraints.

3. Key Contributions

1. Novel Cost Aggregation: Introduction of a normalized mean-CVaR certainty equivalent that balances average performance and tail risk without rescaling travel time units, preserving the convexity of the Beckmann potential.
2. Equivalence Conditions: Proof that path-based and potential-based risk-averse formulations can induce identical equilibrium flows under specific regime-ordering conditions, justifying the use of the more tractable potential-based approach.
3. Distributional Robustness: Development of a TSUE-DRO framework using 1-Wasserstein balls to hedge against distributional misspecification and sampling errors, including extensions for non-stationary, regime-dependent environments.
4. Algorithmic Efficiency: Derivation of a Benders decomposition algorithm specifically tailored for the TSUE-DRO problem, demonstrating superior computational performance compared to standard conic solvers (SCS) on large-scale instances.

4. Results

The study validates the framework through numerical experiments on a stylized 3-link network and a case study of downtown Chicago.

• Small Network (3 Links):

  • Risk Sensitivity: Increasing risk aversion ($\lambda$) or tail sensitivity ($\alpha$) successfully shifts traffic away from links that are fast on average but prone to flooding (Link 1) toward more reliable, albeit slightly slower, alternatives (Links 2 and 3).
  • Truncation: Pure CVaR approaches can be overly conservative, truncating too many routes. The normalized mean-CVaR approach provides a smoother, more realistic trade-off.
  • DRO Impact: The TSUE-DRO formulation yields higher generalized costs and more conservative flow distributions than TSUE-SP, effectively protecting against worst-case distributional shifts.

• Chicago Case Study (3x3 Grid):

  • Flow Redistribution: Under hazard conditions, the model redistributes flows from high-risk bypasses to reliable radial corridors.
  • Quantitative Impact: Compared to a baseline, western corridor traffic increased by 67.9% under TSUE-SP and 100.9% under TSUE-DRO.
  • Stability: The DRO formulation significantly reduced sensitivity to parameter perturbations. For example, under DRO, flow on a critical link increased by ~45% compared to SP, while flow on a risky bypass decreased by ~17%, demonstrating a more robust equilibrium.
  • Computation: The Benders decomposition algorithm converged within ~437 iterations and was roughly 2x faster than standard SOCP solvers for larger network sizes.

5. Significance

• Resilient Planning: The framework provides transportation planners with a tool to design networks that are resilient to extreme weather, moving beyond average-case analysis to account for catastrophic tail events.
• Behavioral Realism: By incorporating tail-risk aversion and ambiguity, the model better reflects how travelers actually behave during hazards (preferring reliability over speed).
• Computational Tractability: The paper bridges the gap between complex risk-averse theory and practical application by proving convexity and providing efficient solution algorithms (Benders/SOCP), making it feasible to apply these models to real-world urban networks.
• Policy Implications: The results suggest that ignoring distributional ambiguity leads to underestimating the need for flow diversification. Robust optimization ensures that traffic assignments remain stable even when hazard probabilities or intensities deviate from historical data.