A Noise-Robust Elicit-to-Optimize Framework for Distortion Riskmetrics via Inverse Reinforcement Learning

Yang Liu; Yuhao Liu; Yunran Wei

REVIEW 2 major objections 4 minor 42 references

A Noise-Robust Elicit-to-Optimize Framework for Distortion Riskmetrics via Inverse Reinforcement Learning

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A noise-robust elicit-to-optimize framework can recover an agent's distortion riskmetric from noisy binary choices at exponential rate, and then use that metric to train policies under a wide class of non-convex risk objectives.

desk verdict The IRL half is a solid, citable contribution; the RL half, as written, does not actually optimize the risk objective it claims—the terminal reward is a state-dependent constant, not a path-dependent risk metric. read the letter →

arxiv 2607.14373 v1 pith:26WK7BJT submitted 2026-07-15 cs.LG q-fin.RM

Yang Liu , Yuhao Liu , Yunran Wei This is my paper · ORCID

classification cs.LGq-fin.RM MSC 68T0591B30

keywords inversereinforcementlearningrisk-sensitivedistortionriskmetricsconditionalriskmeasuresquantileneuralnetworksPPOBayesianelicitationbinarychoicequestions

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish a full pipeline: first, from an agent's noisy answers to simple binary questions, a Bayesian inverse-reinforcement-learning procedure recovers which distortion riskmetric—a broad class of risk measures covering VaR, CVaR, range value-at-risk, and non-convex mixtures—the agent is using, at a convergence rate exponential in the number of questions. Second, the recovered riskmetric is plugged into a model-free reinforcement learning algorithm, an extension of PPO with a quantile network that approximates the conditional cost quantile function, so that policies can be optimized under non-convex, non-coherent risk objectives. The paper proves the existence of finite distinguishing question sets and gives a convergence-rate theorem, and demonstrates both halves on S&P 500 trading tasks. If correct, the framework would unify preference elicitation and risk-sensitive control for a much wider class of risk measures than existing convex/coherent approaches.

What carries the argument

The signed Choquet integral rho_h(X) = ∫ X d(h◦P) with distortion function h, and its quantile representation ∫_0^1 F^{-1}_{X|C}(1-u) dh(u), which converts any distortion riskmetric into an integral of the loss quantile function. This identity is what lets the same object serve both halves: the regret of an action under h defines the pseudo-likelihood for Bayesian elicitation, and the quantile integral defines the terminal reward for the RL optimizer. The distinguishing-question construction (Theorem 1) and the midpoint Riemann sum (Eq. 26) are the concrete mechanisms.

What would settle it

Run Algorithm 2 on a two-action MDP with a known optimal policy under a distortion riskmetric, with the shaping reward (Eq. 16) disabled. If the terminal reward is identical for all trajectories sharing the same initial state, the PPO surrogate has zero path-dependent gradient from the riskmetric, and the learned policy will not change from initialization, contradicting the claim that the algorithm optimizes the riskmetric.

Watch

Extended reading notes

Core claim

The central claim is that an agent's latent risk preference, modeled as a distortion function h (a bounded-variation function on [0,1] defining a signed Choquet integral), can be identified from stochastic binary choices at an exponential convergence rate, and that the identified metric can be used as the objective of a PPO-based policy optimizer. For elicitation, the paper constructs a finite set of binary questions that separates any pair of non-proportional distortion functions (Propositions 1–2, Theorem 1), and a Bayesian pseudo-likelihood scheme (Algorithm 1) whose posterior concentrates on the true h with rate O(exp(-c m + O(sqrt(m log m)))) (Theorem 2), even when the agent chooses sub

Load-bearing premise

The optimization half's load-bearing premise is that setting the terminal reward to the estimated riskmetric value (which depends only on the initial state through the quantile network) and adding a loss-oriented shaping reward yields a policy-gradient signal that actually minimizes the conditional distortion riskmetric; the paper provides no derivation connecting the gradient of the estimated riskmetric to the gradient of the true risk objective.

Editorial extensions

If this is right

An agent's risk preference can be recovered from a few dozen to a hundred noisy binary choices, even when the true risk function is not in the candidate set (the algorithm converges to the closest representative).
Risk-sensitive RL can be applied to a broad class of risk objectives—VaR, CVaR, range value-at-risk, inter-quantile range, inter-ES range, GlueVaR—without needing convexity or coherence.
The quantile-network representation unifies all distortion riskmetrics into a single objective evaluation, so switching the distortion function changes the reward but not the learning algorithm.
Policies trained under different distortion riskmetrics exhibit distinct, interpretable trading boundaries (e.g., CVaR_0.95 being more conservative than CVaR_0.90), suggesting that the elicited metric meaningfully shapes behavior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The optimization half's theoretical gap—no proof that the PPO surrogate's gradient equals the gradient of the riskmetric—suggests the empirical results may be driven by the shaping reward (Eq. 16) rather than by the quantile-based terminal reward; a reader should treat the 'optimize' claim as heuristic until a gradient equivalence is established.
Because the terminal riskmetric is computed from the initial state, the approach is essentially optimizing a state-dependent risk functional of the terminal wealth conditional on the initial condition; a natural extension would be to make the riskmetric path-dependent (e.g., conditioning on the whole trajectory) or to derive a proper policy-gradient theorem for distortion riskmetrics.
An adaptive-refinement variant that grows the candidate set around the current posterior mode would likely improve precision beyond a fixed finite set; the paper's misspecification experiments suggest this is feasible.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

You should know two things about this paper. The IRL section is a genuine contribution: the exponential posterior-convergence rate for noisy binary choices over a finite candidate set of distortion functions is new, and the proof chain (Lemma 2, Prop 4, Cor 1, Theorem 2) is coherent under Assumption 1. Proposition 2 is a clean separating-functional argument, and the elicitation experiments are thorough, including misspecification cases.

The problem is the RL half. The stress-test concern checks out: the paper claims to optimize policies under conditional distortion riskmetrics by extending PPO with a quantile network, but that claim is not supported as written. In Algorithm 2, the terminal reward is set to Ĉ_h, the quantile-network estimate of the risk metric at the initial state s0. For a fixed s0, Ĉ_h is identical for every trajectory, so the PPO surrogate (19) sees no path-dependent gradient from the risk metric. The only trajectory-dependent reward is the ad hoc shaping term (16). No theorem or derivation shows that maximizing expected shaping rewards plus a state-dependent baseline is equivalent to minimizing ρ_h. The experiments in Section 5.2 report only policy heatmaps, with no objective values or comparisons to baselines, so they cannot confirm that ρ_h is actually being minimized.

This is load-bearing because the title and the 'elicit-to-optimize' claim depend on both halves. The IRL part could stand alone as a paper on noise-robust preference elicitation. The RL part needs at least a formal statement of what the PPO update actually optimizes, or a rigorous empirical demonstration with objective values and baselines, before the framework claim holds.

A softer concern: Assumption 1 requires strict inequality ρ_{h_i}(X) ≠ ρ_{h_i}(Y) for all candidates and all questions in the pool. The paper proves a finite distinguishing set exists for each pair, but does not show that the assembled pool avoids ties for every candidate. That is a minor gap, not a fatal one.

Overall, the IRL side is worth a serious referee; the RL side, as presented, is not yet credible. I would send this to review with a clear request to separate the contributions and either fix or downscale the optimization claim. The paper shows serious thinking and honest engagement with the literature, but the central claim is currently overreaching.

Recommendation: reject in current form, but the IRL result is worth a resubmission on its own.

Referee Report

2 major / 4 minor

Summary. The paper proposes a two-stage 'elicit-to-optimize' framework for distortion riskmetrics. In the elicitation stage, the agent's risk preference is assumed to belong to a finite candidate set of distortion functions; the method poses binary-choice questions, models noisy choices with a Boltzmann or a general stochastic choice model, and updates a Bayesian-style pseudo-likelihood. The authors prove finite identifiability (Theorem 1) and an exponential convergence rate for the posterior probability (Theorem 2). In the optimization stage, the paper formulates a conditional distortion riskmetric objective for terminal wealth, represents it as an integral of the quantile function, and extends PPO with policy, value, and quantile networks, setting the terminal reward to the estimated riskmetric (Eq. (26), Algorithm 2).

Significance. If the optimization claim were correct, the paper would unify preference elicitation and risk-sensitive control for a broad class of non-convex risk measures, which is a valuable contribution. The IRL half is well structured: Proposition 2's separation argument, Theorem 1's finite-cover construction, and the martingale-based proof of Theorem 2 are coherent, and the experiments demonstrate robustness to noise and model misspecification. However, the RL half is not supported by a valid derivation: the terminal-reward substitution used in Algorithm 2 does not yield a policy gradient for the distortion riskmetric. Since the central 'elicit-to-optimize' claim depends on both halves, the contribution is not established as it stands.

major comments (2)

[§4.2–4.3, Eq. (26), Algorithm 2 (lines 13–14)] The terminal reward is set to R_T^(n) = Ĉ_h(s_0^(n), φ), a scalar computed by the quantile network at the initial state. For a fixed initial state, this value is identical for every sampled trajectory; it does not depend on the realized path of W_T^π. Consequently, in the PPO surrogate (19) the terminal-reward term is constant as a function of θ, so ∇_θ E[Ĉ_h] = 0. The only trajectory-dependent reward is the ad hoc shaping term (16). No theorem or derivation in the paper establishes that maximizing expected shaping rewards plus a state-dependent baseline is equivalent to minimizing ρ_{h|y0}(W_0 − W_T^π). The numerical section (5.2) reports only policy heatmaps (Figures 10–11), with no objective values or baseline comparisons, so it cannot confirm that the riskmetric is actually minimized. This is a load-bearing gap in the 'optimize' half of the framework.
[§3.2, Assumption 1 and Theorem 1] Theorem 1 constructs, for each candidate h_i and each h ∈ K_{δ,i}, a pair (X_h, Y_h) that separates h_i from h. However, it does not show that the resulting finite pool K satisfies Assumption 1 simultaneously for every h_i ∈ Ĥ and every (X,Y) ∈ K. Assumption 1 is a global no-tie condition, stronger than pairwise distinguishability, and it is used in Lemma 2, Proposition 4, and Theorem 2. The paper never proves that a finite question pool satisfying this simultaneous condition exists. A generic perturbation argument might repair this, but it is not provided. As written, the convergence guarantee is conditional on an unproven structural premise.

minor comments (4)

[§5.1 / Example 1] Example 1 refers to Tables 1–2, but the tables are not present in the manuscript text.
[Figures 1–9] The figures in the submitted PDF appear as placeholder glyphs rather than embedded images; the actual posterior plots should be included.
[§5.2] The RL experiments report only learned policy heatmaps. Training curves and out-of-sample riskmetric values should be reported so the reader can verify that the different objectives lead to measurably different optimized values.
[§3.1 / Equation (12)] The notation Ĥ = {h_1(t), ..., h_L(t)} is slightly informal; Ĥ is a set of functions, not functions evaluated at t. Also, the role of κ in the pseudo-likelihood (learning rate vs. rationality parameter) could be clarified in Remark 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the RL half's reward-to-objective step is an unproven surrogate (a correctness risk), not a circular reduction.

full rationale

The IRL half is self-contained: Theorem 2 is derived from the stated behavioral model (9)-(10), the pseudo-likelihood update (12)-(14), Assumption 1, and martingale concentration arguments; the 'preferred' riskmetric is defined behaviorally and posterior concentration is proven, not assumed. Proposition 2 and Theorem 1 independently construct distinguishing questions with explicit proofs, and the convergence-rate analysis does not presuppose the conclusion. The only self-citation with author overlap is Wang, Wang and Wei (2020), used for the definition of distortion riskmetrics and the quantile representation in Lemma 1; this is an external, published, parameter-free mathematical result, so it counts as independent support rather than a circularity. In the RL half, there is a genuine gap: Eq. (26) approximates rho_h using the quantile network at the initial state, and Algorithm 2 sets the terminal reward R_T to C_hat_h, which is trajectory-independent for a fixed s0; thus the PPO terminal-reward gradient is zero and no theorem shows that the shaping reward (16) optimizes rho_h. However, this is an unverified surrogate/correctness issue, not a case where a prediction is equivalent to its input by construction, and no equation in the paper reduces the central claim to a fitted parameter or a self-citation chain. Therefore, no circular step is exhibited.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central claims rest on the agent model, Assumption 1, the finite-candidate restriction, and the unproven reward substitution for RL. The IRL proof itself is self-contained given these; the RL proof is the largest uncharged premise.

free parameters (6)

κ (pseudo-likelihood learning rate) = 0.2 in experiments; κ>0 in theory
Equation (12); Remark 1 chooses κ=log(0.55/0.45)≈0.2; convergence only needs κ>0 but rates depend on κ.
ε (exploration rate) = 0.01 (adaptive); 1 (random)
Algorithm 1; Corollary 1 uses π0=2ε/[L(L−1)], so exploration is required; no sensitivity analysis.
β (specific decision model rationality) = 0.7, 8 in simulations
Section 3.3.1; controls noise in simulated agents; not estimated by the algorithm.
Pbar / p̂ (general model minimal competence) = 0.6, 0.9 in simulations
Section 3.3.2 and §5.1; assumed lower bound; theorem's rate depends on it.
RL loss coefficients = c0=0,c1=0.04,c2=0.08,c3=0.05,δ=0.05,γ=1,λ_GAE=0.95
Equations (23),(25),(18); chosen by hand; no ablations.
Riemann partition size n for riskmetric integral = unspecified
Equation (26) approximates ∫ quantile dh by n subintervals; error controlled by n but value never given.

assumptions (7)

standard math Quantile representation of distortion riskmetrics (Lemma 1)
Borrowed from Wang et al. (2020); used to define L_h in Proposition 2 and the RL objective; assumed without proof.
domain assumption Atomless probability space with X⊃L∞
Section 2 setup; guarantees existence of quantile functions and boundedness used in Proposition 1 and 3.
domain assumption Assumption 1: no ties for any candidate on any question in K
Stated in §3.2; used in Lemma 2, Prop 4, Theorem 2; realizability of a finite K satisfying it is not proven.
domain assumption Agent is better than random: Pbar>1/2 in general decision model
Equations (9)–(10); Theorem 2's exponential rate has exponent g0=(2Pbar−1)κ, which vanishes if Pbar=1/2.
domain assumption True distortion function lies in finite candidate set Hhat
Convergence theorem is stated for ĥ_l∈Hhat; misspecification experiments are heuristic only.
ad hoc to paper Terminal-reward substitution R_T=−Ĉ_h yields a valid policy gradient for ρ_h
Algorithm 2 lines 12–14; no derivation given; central premise of the RL half.
ad hoc to paper Finite question pool K can be constructed satisfying all assumptions
Theorem 1 only proves existence of pair-specific distinguishing questions; the paper asserts a single pool K without showing it avoids ties for all candidates.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Noise-Robust Elicit-to-Optimize Framework for Distortion Riskmetrics via Inverse Reinforcement Learning." pith.science (2026). https://pith.science/paper/26WK7BJT

@misc{pith2026260714373,
  author       = {Pith},
  title        = {Pith review of: A Noise-Robust Elicit-to-Optimize Framework for Distortion Riskmetrics via Inverse Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/26WK7BJT}},
  note         = {Machine review of arXiv:2607.14373}
}

abstract

We propose a noise-robust elicit-to-optimize framework that integrates inverse reinforcement learning (IRL) and reinforcement learning (RL) for eliciting agents' risk preferences and optimizing policies under a broad class of risk objectives characterized by distortion riskmetrics. On the elicitation side, we propose an adaptive Bayesian IRL method that infers agents' latent risk objectives from their noisy observed decisions, explicitly allowing agents to take stochastic and suboptimal actions. We establish the existence of a finite set of distinguishing questions that identifies the preferred distortion riskmetric within the candidate class and prove that the convergence rate of the algorithm is of order $O(\exp(-cm+O(\sqrt{m\log m})))$ under general settings, where $c>0$ is a constant and $m$ denotes the number of algorithm iterations. On the optimization side, we develop a model-free RL algorithm for optimizing policies under conditional distortion riskmetrics. By representing the objective as an integral of the conditional cost quantile function with respect to the distortion function, the method unifies distortion-riskmetric objectives. We optimize diverse risk objectives by extending the Proximal Policy Optimization (PPO) algorithm with policy, value, and quantile neural networks, where the quantile network estimates the full conditional cost quantile function and enables numerical evaluation of general risk objectives. A comprehensive empirical study demonstrates the framework's elicitation accuracy and effectiveness in complex financial environments.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

[1]

2011 , institution=

CoVaR , author=. 2011 , institution=

2011
[2]

Insurance: Mathematics and Economics , volume=

Parametric measures of variability induced by risk measures , author=. Insurance: Mathematics and Economics , volume=. 2022 , publisher=

2022
[3]

Finance and Stochastics , volume=

A framework for measures of risk under uncertainty , author=. Finance and Stochastics , volume=. 2024 , publisher=

2024
[4]

Insurance: Mathematics and Economics , volume=

Systemic risk: Conditional distortion risk measures , author=. Insurance: Mathematics and Economics , volume=. 2022 , publisher=

2022
[5]

ASTIN Bulletin: The Journal of the IAA , volume=

Distortion riskmetrics on general spaces , author=. ASTIN Bulletin: The Journal of the IAA , volume=. 2020 , publisher=

2020
[6]

Finance and stochastics , volume=

Scenario-based risk evaluation , author=. Finance and stochastics , volume=. 2021 , publisher=

2021
[7]

arXiv preprint arXiv:2411.09659 , year=

A risk sensitive contract-unified reinforcement learning approach for option hedging , author=. arXiv preprint arXiv:2411.09659 , year=

arXiv
[8]

Statistics & Risk Modeling , year =

On dependence consistency of CoVaRand some other systemic risk measures , author =. Statistics & Risk Modeling , year =

Show all 42 references

[9]

The review of financial studies , volume=

Measuring systemic risk , author=. The review of financial studies , volume=. 2017 , publisher=

2017
[10]

The Review of Financial Studies , volume=

SRISK: A conditional capital shortfall measure of systemic risk , author=. The Review of Financial Studies , volume=. 2017 , publisher=

2017
[11]

Stochastic Processes and their Applications , volume=

Risk-consistent conditional systemic risk measures , author=. Stochastic Processes and their Applications , volume=. 2016 , publisher=

2016
[12]

2016 , publisher=

Tobias, Adrian and Brunnermeier, Markus K , journal=. 2016 , publisher=

2016
[13]

Quantitative Finance , volume=

Deep hedging , author=. Quantitative Finance , volume=. 2019 , publisher=

2019
[14]

Mathematical Finance , volume=

Reinforcement learning with dynamic convex risk measures , author=. Mathematical Finance , volume=. 2024 , publisher=

2024
[15]

Journal of Machine Learning Research , volume=

Risk-constrained reinforcement learning with percentile risk criteria , author=. Journal of Machine Learning Research , volume=
[16]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Mean-variance policy iteration for risk-averse reinforcement learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[17]

arXiv preprint arXiv:2505.04553 , year=

Risk-sensitive Reinforcement Learning Based on Convex Scoring Functions , author=. arXiv preprint arXiv:2505.04553 , year=

arXiv
[18]

arXiv preprint arXiv:2403.06323 , year=

A reductions approach to risk-sensitive reinforcement learning with optimized certainty equivalents , author=. arXiv preprint arXiv:2403.06323 , year=

arXiv
[19]

arXiv preprint arXiv:2308.08427 , year=

Eliciting risk aversion with inverse reinforcement learning via interactive questioning , author=. arXiv preprint arXiv:2308.08427 , year=
[20]

Mathematics of Operations Research , volume=

Optimal trend following trading rules , author=. Mathematics of Operations Research , volume=. 2016 , publisher=

2016
[21]

SIAM Journal on Financial Mathematics , volume=

Trend following trading under a regime switching model , author=. SIAM Journal on Financial Mathematics , volume=. 2010 , publisher=

2010
[22]

Journal of Asset Management , volume=

Robo Advisors: quantitative methods inside the robots , author=. Journal of Asset Management , volume=. 2018 , publisher=

2018
[23]

Management Science , volume=

Personalized robo-advising: Enhancing investment through client interaction , author=. Management Science , volume=. 2022 , publisher=

2022
[24]

Artificial Intelligence , volume=

A survey of inverse reinforcement learning: Challenges, methods and progress , author=. Artificial Intelligence , volume=. 2021 , publisher=

2021
[25]

International Conference on Machine Learning , pages=

Towards theoretical understanding of inverse reinforcement learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

2023
[26]

Advances in Neural Information Processing Systems , volume=

Identifiability in inverse reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=
[27]

International Conference on Machine Learning , pages=

Reward identification in inverse reinforcement learning , author=. International Conference on Machine Learning , pages=. 2021 , organization=

2021
[28]

Joint European conference on machine learning and knowledge discovery in databases , pages=

Active learning for reward estimation in inverse reinforcement learning , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2009 , organization=

2009
[29]

International Conference on Machine Learning , pages=

Interactive inverse reinforcement learning for cooperative games , author=. International Conference on Machine Learning , pages=. 2022 , organization=

2022
[30]

arXiv preprint arXiv:1707.06347 , year=

Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

arXiv
[31]

arXiv preprint arXiv:2405.04714 , year=

Racer: Epistemic risk-sensitive rl enables fast driving with fewer crashes , author=. arXiv preprint arXiv:2405.04714 , year=

arXiv
[32]

The Palgrave handbook of technological finance , pages=

Robo-advising , author=. The Palgrave handbook of technological finance , pages=. 2021 , publisher=

2021
[33]

2019 IEEE Intelligent Vehicles Symposium (IV) , pages=

Addressing inherent uncertainty: Risk-sensitive behavior generation for automated driving using distributional reinforcement learning , author=. 2019 IEEE Intelligent Vehicles Symposium (IV) , pages=. 2019 , organization=

2019
[34]

, author=

Risk-sensitive Inverse Reinforcement Learning via Coherent Risk Models. , author=. Robotics: science and systems , volume=
[35]

2020 IEEE Intelligent Vehicles Symposium (IV) , pages=

Risk-aware high-level decisions for automated driving at occluded intersections with reinforcement learning , author=. 2020 IEEE Intelligent Vehicles Symposium (IV) , pages=. 2020 , organization=

2020
[36]

2015 IEEE international conference on robotics and automation (ICRA) , pages=

Learning driving styles for autonomous vehicles from demonstration , author=. 2015 IEEE international conference on robotics and automation (ICRA) , pages=. 2015 , organization=

2015
[37]

, author=

Planning for autonomous cars that leverage effects on human actions. , author=. Robotics: Science and systems , volume=. 2016 , organization=

2016
[38]

Journal of Financial Econometrics , volume=

Robo-advising: Learning investors’ risk preferences via portfolio choices , author=. Journal of Financial Econometrics , volume=. 2021 , publisher=

2021
[39]

Robotics science and systems workshop on inverse optimal control and robotic learning from demonstration , pages=

Inverse optimal control for humanoid locomotion , author=. Robotics science and systems workshop on inverse optimal control and robotic learning from demonstration , pages=
[40]

2010 IEEE International Conference on Robotics and Automation , pages=

An optimization approach to rough terrain locomotion , author=. 2010 IEEE International Conference on Robotics and Automation , pages=. 2010 , organization=

2010
[41]

INFORMS Journal on Computing , volume=

Computing sensitivities for distortion risk measures , author=. INFORMS Journal on Computing , volume=. 2021 , publisher=

2021
[42]

INFORMS Journal on Computing , volume=

Transfer reinforcement learning for mixed observability markov decision processes with time-varying interval-valued parameters and its application in pandemic control , author=. INFORMS Journal on Computing , volume=. 2025 , publisher=

2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

2011 , institution=

CoVaR , author=. 2011 , institution=

2011

[2] [2]

Insurance: Mathematics and Economics , volume=

Parametric measures of variability induced by risk measures , author=. Insurance: Mathematics and Economics , volume=. 2022 , publisher=

2022

[3] [3]

Finance and Stochastics , volume=

A framework for measures of risk under uncertainty , author=. Finance and Stochastics , volume=. 2024 , publisher=

2024

[4] [4]

Insurance: Mathematics and Economics , volume=

Systemic risk: Conditional distortion risk measures , author=. Insurance: Mathematics and Economics , volume=. 2022 , publisher=

2022

[5] [5]

ASTIN Bulletin: The Journal of the IAA , volume=

Distortion riskmetrics on general spaces , author=. ASTIN Bulletin: The Journal of the IAA , volume=. 2020 , publisher=

2020

[6] [6]

Finance and stochastics , volume=

Scenario-based risk evaluation , author=. Finance and stochastics , volume=. 2021 , publisher=

2021

[7] [7]

arXiv preprint arXiv:2411.09659 , year=

A risk sensitive contract-unified reinforcement learning approach for option hedging , author=. arXiv preprint arXiv:2411.09659 , year=

arXiv

[8] [8]

Statistics & Risk Modeling , year =

On dependence consistency of CoVaRand some other systemic risk measures , author =. Statistics & Risk Modeling , year =