[Paper Review] Risk-Adaptive Approaches to Stochastic Optimization: A Survey

Johannes Ø. Røyset|arXiv (Cornell University)|Dec 1, 2022

Risk and Portfolio Optimization4 citations

TL;DR

This survey presents a comprehensive overview of risk-adaptive approaches in stochastic optimization, emphasizing risk measures rooted in convex analysis. It unifies diverse applications in engineering, machine learning, and finance by integrating risk-averse decision-making, robust optimization, and distributionally robust frameworks, with key contributions in theoretical foundations, algorithms, and emerging applications such as fair machine learning and reliability measures.

ABSTRACT

Uncertainty is prevalent in engineering design, data-driven problems, and decision making broadly. Due to inherent risk-averseness and ambiguity about assumptions, it is common to address uncertainty by formulating and solving conservative optimization models expressed using measures of risk and related concepts. We survey the rapid development of risk measures over the last quarter century. From their beginning in financial engineering, we recount the spread to nearly all areas of engineering and applied mathematics. Solidly rooted in convex analysis, risk measures furnish a general framework for handling uncertainty with significant computational and theoretical advantages. We describe the key facts, list several concrete algorithms, and provide an extensive list of references for further reading. The survey recalls connections with utility theory and distributionally robust optimization, points to emerging applications areas such as fair machine learning, and defines measures of reliability.

Motivation & Objective

• To unify and systematize the development of risk measures in stochastic optimization over the past 25 years.
• To provide a theoretical and computational framework for risk-averse decision-making under uncertainty in engineering, statistics, and machine learning.
• To connect risk measures with utility theory, distributionally robust optimization, and emerging areas like fairness in machine learning.
• To identify open problems and challenges in dynamic, multi-stage, and partially observable optimization under risk.
• To introduce and formalize measures of reliability as alternatives to traditional failure probabilities.

Proposed method

• The paper employs convex analysis to formalize risk measures, including superquantiles (conditional value-at-risk) and deviation measures, as core tools for conservative optimization.
• It introduces risk envelopes and duality theory to derive dual algorithms and connect risk measures with distributionally robust optimization.
• The framework incorporates regret and error measures to model decision performance under uncertainty, with applications in regression and machine learning.
• It formulates risk-averse optimization problems using law-invariant and coherent risk measures, enabling robustness to distributional ambiguity.
• The paper proposes algorithms for superquantile minimization and estimation, leveraging convex conjugacy and subgradient methods.
• It extends the framework to dynamic and multi-stage settings, and discusses risk-averse experimental design and surrogate modeling.

Experimental results

Research questions

• RQ1 How can risk measures be systematically applied to unify robust and stochastic optimization across engineering, statistics, and finance?
• RQ2 What are the theoretical and computational advantages of using superquantiles and risk envelopes in optimization under uncertainty?
• RQ3 How can risk-averse formulations improve fairness and reliability in machine learning and decision-making systems?
• RQ4 In what ways do risk measures connect with utility theory and distributionally robust optimization?
• RQ5 What are the open challenges in extending risk-adaptive methods to dynamic, multi-stage, and partially observable decision processes?

Key findings

• Risk measures, particularly superquantiles and conditional value-at-risk, provide a robust and computationally tractable framework for conservative decision-making under uncertainty.
• The duality theory of risk measures enables efficient algorithms through risk envelopes and subgradient methods, linking to distributionally robust optimization.
• The integration of risk measures with regression and machine learning allows for improved performance tracking and fairness in predictive models.
• Measures of reliability are proposed as alternatives to traditional failure probabilities, especially in safety-critical systems.
• The framework supports risk-averse experimental design and surrogate modeling, with potential to reduce underestimation of critical response variables.
• Despite progress, Bayesian risk-averse optimization and decision-dependent probability measures remain underexplored, highlighting key open challenges.