Source-linked AI summary
Conformal Risk-Averse Decision Making with Optimized Certainty Equivalent Risk Control
Amirmohammad Farzaneh, Osvaldo Simeone
TL;DR
Risk-averse decisions must control uncertainty and tail losses rather than optimize expected loss alone. The paper develops an OCE-based policy framework, derives prediction-set structure for CVaR, and calibrates unknown-distribution policies with held-out data. In wireless beamforming, the CVaR-optimal strategy satisfies the target in all 500 trials, while the VaR baseline exceeds it in 53.4% of trials.
Problem
Decision making under uncertain system states needs principled risk control because expected loss can downplay variability and tail events.
Method
The paper combines OCE risk optimization, CVaR prediction-set characterization, and learn-then-test calibration using a learned conditional model and held-out calibration data.
Results
In the ray-traced beamforming setting, the VaR baseline exceeds target CVaR_0.20 in 53.4% of trials, whereas the CVaR-optimal baseline satisfies the target in all 500 trials.
Takeaways & Limitations
For CVaR, the framework yields a prediction-set policy that minimizes average tail loss outside the covered set rather than maximum loss inside it.
Abstract
from arXiv · showhide
We study risk-averse decision making, in which an agent selects actions while being uncertain about the true system state. The risk is measured via optimized certainty equivalent (OCE) metrics, which generalize popular criteria such as mean-variance risk and conditional value-at-risk (CVaR). We characterize the optimal policy under known distributions, and show that it reduces to a prediction set-based solution for the CVaR. This provides an operational interpretation of conformal prediction-type prediction sets. For unknown distributions, we develop a data-driven calibration strategy, based on a synthetic model for the likelihood and held-out calibration data, yielding high-probability control of the OCE risk. The approach is evaluated on two wireless beamforming settings.
I. INTRODUCTION
The paper frames risk-averse decisions as requiring uncertainty-aware risk control beyond expected loss, and proposes OCE-based policies with conformal prediction-set interpretations and calibration for unknown distributions.
- Risk-averse decisions must account for predictive uncertainty and tail-loss variability, which expected-loss minimization can understate.
- OCE risk measures provide a broad convex framework encompassing CVaR, entropic risk, and mean-variance risk.
- VaR-based conformal decision rules use hard coverage thresholds that do not distinguish moderate from extreme tail losses.
- The paper formulates risk-averse decision making with OCE metrics and derives the optimal policy when distributions are known.
- For unknown distributions, the proposed calibration algorithm uses learn-then-test to provide high-probability OCE-risk control and is evaluated in two wireless beamforming settings.
II. PROBLEM FORMULATION
The formulation models actions chosen from observations with losses depending on an unobserved state, then uses OCE risk to control risk-averse performance through a reserve and penalty function.
- An observation-dependent policy selects actions without observing the true state, incurring a loss determined by the action and state.
- OCE risk decomposes loss into a deterministic reserve and a random residual penalized in expectation, optimizing over the reserve.
- The penalty function determines risk aversion, with OCE including mean-variance risk and CVaR as special cases.
- For continuous losses, CVaR averages losses in the worst α fraction beyond the VaR threshold.
- Smooth CVaR introduces a positive temperature parameter and converges to standard CVaR as the temperature approaches zero.
- The risk-averse policy optimization minimizes the OCE risk associated with a prespecified penalty function.
III. OPTIMAL POLICY FOR KNOWN STATE-OBSERVATION DISTRIBUTION
With a known joint observation-state distribution, the framework characterizes the optimal policy and connects the CVaR case to a prediction-set solution.
- For known distributions, the paper characterizes the solution by optimizing the policy under the OCE objective and recovering a CVaR prediction-set structure.
A. Optimal Policy
For a fixed reserve, the known-distribution optimization separates across observations, yielding conditional action optimization followed by scalar reserve optimization and stationarity-based solution methods.
- Interchanging the objective’s two minimizations makes the optimal action separable across covariates for a fixed reserve.
- The fixed-reserve action minimizes the conditional expected penalized loss over the state distribution given each observation.
- After optimizing conditional actions, the remaining OCE problem reduces to a scalar optimization over the reserve.
- When differentiable, the optimal reserve satisfies a stationarity condition involving the expected derivative of the penalty.
- The reserve condition can be approximately solved by repeatedly optimizing actions for fixed reserves and applying a root-finding method such as a Chebyshev proxy.
B. Optimal Policy via Set Prediction
The CVaR formulation yields a prediction-set-based policy, but unlike VaR's max-min rule, it optimizes average tail loss beyond the reserve. The resulting reserve equals the loss VaR and is more conservative than the feature-dependent VaR formulation.
- The prior VaR formulation jointly chooses actions and feature-dependent thresholds using a max-min rule over prediction sets with marginal coverage at least 1 −α.
- The OCE-CVaR formulation complements VaR by targeting average loss over the worst α-fraction rather than guaranteed loss over the best 1 −α-fraction.
- As the smoothing parameter τ approaches zero, the CVaR solution induces a prediction set C∗(x) with marginal coverage 1 −α.
- The CVaR-optimal action minimizes expected excess loss above reserve t∗ conditional on the observation, rather than maximum loss inside the covered set.
- The reserve t∗ equals the loss VaR and satisfies t∗ ≥ E[t∗(X)], reflecting the greater conservatism of the CVaR objective.
IV. DATA-DRIVEN CALIBRATION
For unknown distributions, the method combines a conditional likelihood model with held-out calibration over candidate reserves. An upper-confidence-bound selection rule then certifies the deployed policy's CVaR risk with high probability.
- The calibration setting uses a pretrained conditional model ˆP_Y|X and an independent held-out calibration set to select a reserve and control CVaR risk.
- For each candidate reserve, the procedure computes a model-based action by minimizing expected penalty under the modeled conditional state distribution.
- The model-based policy uses only ˆP_Y|X and is generally suboptimal for the true distribution P_XY before calibration.
- Calibration evaluates realized penalties on true labels, whose empirical average upper-bounds the OCE risk at a fixed reserve.
- A finite reserve grid, bounded penalties, Hoeffding's inequality, and a union bound produce a simultaneous high-probability upper bound over candidate reserves.
- The algorithm selects the smallest candidate reserve whose upper confidence bound is at most ε and outputs the corresponding policy with the stated guarantee.
V. NUMERICAL RESULTS
The paper evaluates its optimal policy and data-driven calibration scheme on wireless beamforming under known and unknown channel distributions.
- The experiments cover line-of-sight beamforming with known PXY and ray-traced beamforming with unknown PXY.Both settings use an eight-antenna base station, a single-antenna user, normalized beamforming vectors, and normalized capacity-gap loss.
A. Line-of-Sight Channel
In the line-of-sight setting, CVaR-based optimization avoids the abrupt VaR decision shift and achieves decreasing CVaR as positive-error probability increases.
- The experiment models angle estimates on [−30°, +30°] with bimodal ±3° errors and beams selected from angular offsets θ∈[−8°, +8°].The channel includes array response and complex Gaussian noise with σh = 0.2.
- Fig. 1 compares the optimized offset at X = 0 and empirical CVaRα for VaR-based and CVaR-minimizing policies.The CVaR policy uses an sCVaR penalty with τ = 10^-2 and solves equation (9) via a Chebyshev proxy method.
- For positive-error probabilities in [0.5, 0.7), both criteria choose a compromise beam near θ = 0°, producing similar CVaR0.30.
- At probability 1−α, the VaR baseline abruptly selects θ = +3°, whereas the CVaR-optimal strategy gradually accommodates negative errors and yields decreasing CVaR.The VaR shift occurs because outcomes with probability below 0.3 do not affect VaR optimization at α = 0.3.
B. Ray-Traced Channel
In the ray-traced setting with unknown distributions, the calibrated CVaR-optimal scheme satisfies the target in every trial, while the VaR baseline frequently exceeds it.
- The experiment uses Sionna ray tracing of the Munich scene with estimated UE location and bimodal additive location error of magnitude 6 m.
- 53.4% of the VaR baseline trials exceeded ε, violating the guarantee, while the CVaR-optimal scheme met CVaR0.20 < ε in all 500 trials.The empirical distribution is computed over 500 Monte Carlo redraws of test channel realizations.
VI. CONCLUSIONS
The paper unifies OCE-based risk-averse decision making, deriving prediction-set structure for CVaR and a calibrated procedure for high-probability risk control.
- For known state–observation distributions, the optimal policy uses conditional action optimization together with scalar optimization over a reserve.
- For CVaR, the known-distribution characterization yields a prediction-set interpretation complementary to VaR-based decision making.
- For unknown distributions, calibration combines a learned conditional model with held-out data to provide high-probability control of OCE risk.
- Numerical experiments on wireless beamforming confirmed the theoretical findings, while future work may examine larger-scale experiments and other OCE measures.