Additive consistency of risk measures and its application to risk-averse\n routing in networks

Roberto Cominetti, Alfredo Torrico

Abstract

This paper investigates the use of risk measures and theories of choice for modeling risk-averse route choice and traffic network equilibrium with random travel times. We interpret the postulates of these theories in the context of routing, and we identify additive consistency as a plausible and relevant condition that allows to reduce risk-averse route choice to a standard shortest path problem. Within the classical theories of choice under risk, we show that the only preferences that satisfy this consistency property are the ones induced by the entropic risk measures.

Introduction

When travel times in a transportation network are stochastic, drivers often exhibit risk-averse behavior, preferring reliable routes over those with high variance, even if the latter have lower expected travel times. However, modeling this behavior mathematically presents significant challenges. Many common risk models, such as the mean-standard deviation approach, can lead to "paradoxical" decisions where the addition of a common, independent segment to two alternative paths can flip a driver's preference.

A network illustrating the choice between paths X and Y followed by a common segment Z.

Consider the network in Figure 1. A driver chooses between route $X$ and route $Y$ to reach junction $j$, and then must take segment $Z$ to reach destination $d$. Additive consistency requires that if a driver prefers $X$ over $Y$, they must also prefer the combined path $X+Z$ over $Y+Z$, provided $Z$ is independent of $X$ and $Y$. This property is essential for ensuring that optimal paths are composed of optimal sub-paths, a prerequisite for efficient computation in large-scale networks.

Theoretical Framework

The authors evaluate three major economic theories of choice under risk to identify which measures satisfy additive consistency:

  1. Expected Utility Theory (EUT): Minimizing the expected value of a disutility function $u(X)$.
  2. Dual Theory of Choice: Utilizing distortion risk measures where probabilities are weighted (distorted) rather than costs.
  3. Rank-Dependent Expected Utility (RDEU): A hybrid model combining both cost and probability distortions.

The core of the analysis involves testing these frameworks for the property $\rho(X+Y) = \rho(X) + \rho(Y)$ for independent random variables $X$ and $Y$.

The Entropic Monopoly and Translation Invariance

A key finding is the unique role of the entropic risk measure, defined as $\rho(X) = \frac{1}{\beta} \ln(E[e^{\beta X}])$. This measure is shown to be the only one within the EUT framework that satisfies translation invariance.

A network illustrating translation invariance where the common segment Z is a constant delay m.

Translation invariance, illustrated in Figure 2, implies that adding a deterministic delay $m$ to any path increases its perceived risk by exactly $m$. The paper proves that for Expected Utility maps, translation invariance is equivalent to being an entropic risk measure (Theorem 3.1). Furthermore, the authors demonstrate that in the Dual Theory of Choice, the only additively consistent measure is the risk-neutral expected value (Theorem 3.2). Even in the more general RDEU model, the entropic measure remains the sole consistent choice for risk-averse behavior (Theorem 3.3).

Implications for Network Modeling

The discovery that entropic risk measures are uniquely additively consistent has profound practical implications for transportation engineering:

  • Shortest Path Reduction: Because the risk of a path is the sum of the risks of its links, finding the most preferred risk-averse route reduces to a standard shortest-path problem. This allows for the use of efficient algorithms like Dijkstra's.
  • Equilibrium Stability: These measures integrate seamlessly into classical traffic equilibrium models, such as Wardrop equilibrium. This enables the prediction of large-scale traffic patterns under risk-averse behavior using existing computational tools.
  • Critique of Non-Additive Measures: The paper highlights the failure of the popular "Mean + Standard Deviation" model in network contexts. Since the standard deviation of a sum of independent variables is not the sum of their standard deviations, this model violates additive consistency and cannot be solved using standard link-based shortest path algorithms.

Conclusion and Future Directions

This research establishes a rigorous mathematical foundation for risk-averse routing, identifying entropic risk measures as the only choice within classical frameworks that maintain consistency across network segments.

While theoretically robust, the model has limitations. It assumes link independence, which may not hold in congested urban networks where delays are correlated. Additionally, it assumes a homogeneous population of drivers with identical risk-aversion parameters ($\beta$). Future work should explore multi-user equilibrium models and the impact of correlated travel times on these consistency properties. Finally, the gap between mathematical "rationality" (additive consistency) and actual human behavioral patterns remains a fertile ground for future interdisciplinary study.

Summary updated

Abstract

This paper investigates the use of risk measures and theories of choice for\nmodeling risk-averse route choice and traffic network equilibrium with random\ntravel times. We interpret the postulates of these theories in the context of\nrouting, and we identify additive consistency as a plausible and relevant\ncondition that allows to reduce risk-averse route choice to a standard shortest\npath problem. Within the classical theories of choice under risk, we show that\nthe only preferences that satisfy this consistency property are the ones\ninduced by the entropic risk measures.\n
BibTeX
@article{cominetti2016additive,
  author = {Roberto Cominetti and Alfredo Torrico},
  title = {{Additive consistency of risk measures and its application to risk-averse\textbackslash\{\}n routing in networks}},
  journal = {Math. Oper. Res.},
  year = {2016},
  doi = {10.1287/moor.2016.0787},
}

Loading related content...

Figure