Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Information-Theoretic Aggregation of Ethical Attributes in Simulated-Command

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that ethical-attribute weights in simulated command decisions can be derived automatically from each attribute's variance across situations, moving human judgment outside the simulation loop.

desk verdict The paper's central ICW formula reverses its own stated principle, and IGD is mathematically degenerate; this is not sound as submitted, though the framing of separating variance information from mean information is a useful seed. read the letter →

arxiv 2507.12862 v2 pith:L7ZNPN3Z submitted 2025-07-17 cs.AI

Taylan Akay , Harrison Tolley , Hussein Abbass This is my paper · ORCID
classification cs.AI
keywords ethicalAImulti-criteriadecisionmakingentropyweightingvariance-basedweightssimulation-basedtestandevaluationautonomousweaponsystemsinformationgainKullback-Leiblerdivergence
open problems The Measurement Problem
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 solve the weighting problem inside simulation-based test and evaluation of AI command systems when the attributes being weighted are ethical. The core claim is that an attribute's second central moment—how much its utility varies across situations for the same course of action—carries the information needed to set its weight, so weights are computed from the variance matrix Lambda while decisions are scored with the mean matrix Gamma. Three low-cost information-theoretic formulas are offered: an entropy weight (ICW), a Kullback-Leibler weight against a human prior (IGH), and an entropy-difference weight (IGD). If this is right, simulated environments can explore huge scenario spaces and come back to the human commander with a few ranked options, without a human weighting every in-simulation decision. The paper's working premise, stated openly, is that the ethical measurement problem is solved and the utilities already embed the commander's subjective importance.

What carries the argument

The load-bearing objects are two matrices built from normalised utilities U(m,r,s): Gamma holds first central moments (mean utility of each alternative for each attribute) and Lambda holds second central moments (variance of that utility across situations). From Lambda the paper forms p(m,r), the relative variance of alternative m for attribute r, and then three weight families: ICW uses Shannon entropy of p, IGH uses Kullback-Leibler divergence between p and a human-supplied prior q, and IGD uses the difference between an attribute's entropy and the average entropy of the others. These formulas convert 'how much an attribute's utility wobbles across situations' into a normalised weight, and they are what let a simulation score alternatives without a human in the per-decision loop.

What would settle it

Collect a corpus of simulated command situations where a known ethically decisive attribute is constant across all situations for every course of action; if the proposed weights give that attribute zero weight, an externally validated judgment would show the weighting contradicts what human commanders consider decisive.

Watch

Extended reading notes

Core claim

The central discovery the paper argues for is that the variability of an ethical attribute across situations is itself an information source about how much that attribute should count, and that this information should be separated from the mean performance used to make decisions. Weights are therefore to be derived from Lambda, the matrix of per-alternative, per-attribute variances of normalised utilities, while Gamma, the matrix of mean utilities, is reserved for computing expected utility and ranking courses of action. Three concrete aggregation formulas are proposed: ICW normalises the entropy of each attribute's relative variance distribution; IGH computes the Kullback-Leibler divergence between the simulation-driven variance distribution and a subjective prior supplied by the human; and IGD computes the difference between an attribute's entropy and the average entropy of the other attributes. In the synthetic autonomous-weapon use case, the three methods yield different weights and can reverse which scenario is preferred, which the paper presents as evidence that the choice of information source and prior matters.

Load-bearing premise

The paper assumes the ethical measurement problem is solved: acceptable, granular ethical metrics exist and the utility values U(r,s) already carry the commander's subjective importance, so weighting failures are attributed to the weights alone.

Editorial extensions

If this is right

  • Simulation-based test and evaluation can rank courses of action by ethically weighted expected utility without a human setting weights for each scenario, leaving human judgment to metric design before simulation and option choice after it.
  • An ethical attribute whose utility barely varies across situations is automatically downweighted, because it carries little discriminatory information; an attribute that fluctuates strongly in complex situations is upweighted.
  • The IGH variant lets a commander encode a subjective prior on attribute importance before simulation, and the resulting weights are pulled toward that prior only when simulation data diverge from it.
  • Because ICW, IGH, and IGD can reverse the preferred scenario in the paper's use case, the weight-derivation rule itself is part of the ethical assessment and must be chosen deliberately.
  • Using Lambda for weights and Gamma for decisions resolves the known limitation that classic entropy weights punish variability in alternatives rather than rewarding discriminatory power across situations.

Reading between the lines

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

  • Beyond the paper, the same variance-as-weight principle could be applied to any intangible attribute in simulation-based evaluation—risk, trust, or mission impact—not only ethical attributes, wherever a defensible utility mapping exists.
  • Beyond the paper, the method suggests a testable calibration loop: collect human-assigned weights for a corpus of simulated command decisions and compare them with ICW/IGD weights to see whether variance tracks what commanders actually treat as informative.
  • Beyond the paper, the three formulas could be embedded inside an autonomous system as an internal deliberation step, so an agent re-derives its ethical trade-off weights on the fly; the paper mentions this possibility but does not develop it.
  • Beyond the paper, the IGH variant offers a way to monitor value drift: large and persistent divergence between the human prior and simulation-driven probabilities would flag that the metric space or the prior no longer matches the situations being generated.
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.

Referee Report

3 major / 4 minor

Summary. This paper proposes to automate the weighting of ethical attributes in simulation-based test and evaluation. It defines a mean matrix Γ and a variance matrix Λ over alternatives and situations, then derives three information-theoretic weighting schemes from Λ: ICW (Shannon entropy of normalized variances), IGH (Kullback-Leibler divergence from a human-supplied prior), and IGD (a difference between an attribute's entropy and a function of the remaining entropies). A two-attribute, two-alternative toy scenario is used to illustrate the weights and the resulting scenario rankings. The authors conclude that the results demonstrate the importance of incorporating human judgment into the decision-making process.

Significance. If the proposed methods were valid, they would address a relevant problem in multi-criteria decision analysis and simulation-based testing: how to aggregate ethical attributes without requiring human input at each simulation step. The paper is transparent about its key assumptions, including that ethical metrics exist and that utilities encode subjective judgment, and it provides a fully worked numerical example. However, the central methodological claims are not established. The primary entropy-based weight formula rewards the opposite of the paper's stated principle; the IGD formula is degenerate and cannot differentiate attributes; and the IGH-based conclusion about human judgment follows by construction rather than from empirical evidence. The paper does not provide machine-checked proofs, reproducibility artifacts, or a comparison to existing entropy-weighting methods, so the positive contribution is limited to a clearly stated problem formulation and a reproducible arithmetic illustration.

major comments (3)
  1. [Section III-A, Section III-C, Table III] The ICW formula is inverted with respect to the paper's own premise. Section III-C defines p(m,r) as the variance of alternative m for attribute r normalized over alternatives, and E(r) as the Shannon entropy of p. Since entropy is maximized by a uniform distribution, an attribute whose normalized variances are closest to being equal across alternatives receives the highest ICW. This is precisely the situation that Section III-A says should receive low weight: 'An attribute should not be weighted high if it has similar variance on all courses of actions.' The toy example confirms the inversion: Force Protection has normalized variances (0.74, 0.26) and Proportionality has (0.2, 0.8); after normalization, Force Protection is closer to the uniform distribution, and Table III assigns it the higher ICW weight (0.54 vs 0.46). Thus, the claim that ICW reflects discriminatory power is not supported by the formula as written.
  2. [Section III-C, Appendix A5, Table III] The IGD method is mathematically degenerate. Since IGD(r) = E(r) - sum_{j != r} (1 - E(j)) = sum_j E(j) - (R - 1), the value is independent of r for any number of attributes R. Consequently, IGDW(r) = 1/R for every attribute whenever the denominator is nonzero, so the method can never differentiate attributes. The appendix confirms this: Table A5 reports identical IGD values of 0.0767 for both attributes, and Table III reports IGDW weights of 0.50 and 0.50. The IGD method as defined is vacuous and should either be removed or redefined with a formula that actually depends on r.
  3. [Section III-C, Section IV-B] The paper's conclusion that 'the results reveal the importance for incorporating human judgement into the decision making process' is circular. IGHW(r) is, by definition, the normalized Kullback-Leibler divergence between the simulation-driven distribution p(.,r) and the human prior q(.,r). An attribute that disagrees with the prior is therefore guaranteed to receive a high weight, and an attribute that agrees with the prior receives a low weight; if p = q, the attribute receives zero weight and is eliminated from the aggregation. The ranking difference between IGHW and the data-only methods in Table III is a direct consequence of this definition, not an empirical finding. To support the conclusion, the authors would need an independent criterion, such as ground-truth ethical outcomes or decision-maker validation, showing that divergence from the prior corresponds to attribute importance.
minor comments (4)
  1. [Section III-A, Table I] Section III-A states that utilities are normalized to U : (r,s) -> [0,1], but Table I reports means of 15, 8, 4, and 10 and variances up to 7.0, which cannot arise from a [0,1]-valued utility. Please rescale the example or explicitly relax the normalization assumption.
  2. [Section III-C, Appendix A3, Appendix A4] The logarithm base is inconsistent: Section III-C uses log base 2 in the ICW definition, while Appendix A3 uses natural logarithms; the IGH definition in Section III-C is written with an unspecified log, and Appendix A4 uses base-10 logarithms. Because the weights are normalized, the final weights are invariant to the base, but the reported intermediate entropy and IGH values do not match the formulas as written.
  3. [Appendix A5] The appendix describes IGD as 'comparing the entropy of each attribute to the average informativeness of the others,' but the formula in Section III-C is E(r) - sum_{j != r} (1 - E(j)), which is neither an average nor a difference with the average entropy of the other attributes. The description and the formula should be aligned.
  4. [Section III-C] The paper does not discuss the degenerate case in which all variances of an attribute are zero; in that case p(m,r) is undefined because the denominator of the normalization is zero. A brief note on handling zero total variance would make the method more robust.

Circularity Check

1 steps flagged · score 4.0 of 10

IGH demonstration restates its definition; ICW/IGD issues are correctness, not circularity.

  1. self definitional [Section III-C (IGH definition); Section IV-B (Results)]
    "In the first form, we call it IGH to emphasise its reliance on human input ... IGH will combine the discriminatory power of an attribute as measured within a simulation environment with the subjective belief of its importance as measured by the decision maker. The further away the simulation-driven probabilities are from the subjective ones, the higher the weight. ... The results reveal the importance for incorporating human judgement into the decision making process."

    IGH weight is the normalized KL divergence between the variance-derived distribution p and a human-supplied prior q, so any discrepancy between p and q automatically raises IGH and changes the ranking. The use-case observation that human judgement matters (IGHW 0.07 vs 0.93; scenario ranking flips) follows from the chosen q differing from p for Proportionality; it holds by construction, not as an empirical finding. The paper's wording 'The results reveal...' presents this defining property as an outcome, making the demonstration self-definitional. The other two methods are data-driven and this circularity does not infect them; ICW's inversion of the stated variance-similarity principle and IGD's reduction to uniform weights are separate correctness failures, not circularity.

full rationale

No load-bearing self-citation or imported uniqueness theorem exists; the weighting formulas are derived in the paper from the variance matrix Lambda rather than from prior work by the same authors. The only genuine circularity is the IGH demonstration, where the claimed 'result' that human priors change weights and ranking is a direct algebraic consequence of defining IGH as KL divergence against a human prior: Tables II and III make this explicit (q for Proportionality is 0.1/0.9 versus p 0.2/0.8, giving IGHW 0.93). This partial self-definitional step warrants score 4, not higher, because the central weighting scheme also includes data-driven ICW and IGD, and the paper's core derivation chain is not itself forced by an input-output equivalence. The ICW and IGD problems identified by the skeptic are substantive correctness concerns (ICW reverses the paper's own 'similar variance should lower weight' premise; IGD algebraically equals sum_j E(j) - (R-1), a constant independent of the attribute), but they are not circularity, so they do not raise the score beyond the noted partial circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's methods rest on the domain assumption that ethical metrics exist and are valid, on the modeling choice that variance encodes importance, and on the standard mathematics of entropy. The only fitted input is the subjective prior in IGH.

free parameters (1)
  • Subjective prior q(i,r) in IGH = Force Protection: (0.7, 0.3); Proportionality: (0.1, 0.9)
    The human-provided prior directly determines the IGH weights via KL divergence; changing it changes the final ranking. No elicitation procedure is given.
assumptions (4)
  • domain assumption Ethical attribute metrics are adequately designed to assess ethical implications
    Stated in the Abstract and Section I: 'We assume that the problem of designing metrics that are sufficiently granular to assess the ethical implications of decisions is solved.'
  • domain assumption Normalized utilities U(r,s) already embed the decision maker's subjective judgment about attribute importance
    Section III-B: 'Our reliance on utilities, U(r,s), and not on measurements, E(r,s), assumes that the decision maker's subjective judgement has been incorporated in the construction of these utilities.'
  • ad hoc to paper The second central moment (variance) of an attribute across situations is a valid proxy for its discriminatory power and importance
    Section III-A: 'Our premise is that the second-central moment is a source of information that could be used for this purpose.' No empirical or theoretical justification is provided beyond an illustrative example.
  • standard math Entropy and Kullback-Leibler divergence are appropriate measures of attribute informativeness
    Standard tools from information theory, invoked in Section III-C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Information-Theoretic Aggregation of Ethical Attributes in Simulated-Command." pith.science (2026). https://pith.science/paper/L7ZNPN3Z

@misc{pith2026250712862,
  author       = {Pith},
  title        = {Pith review of: Information-Theoretic Aggregation of Ethical Attributes in Simulated-Command},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7ZNPN3Z}},
  note         = {Machine review of arXiv:2507.12862}
}
read the original abstract

In the age of AI, human commanders need to use the computational powers available in today's environment to simulate a very large number of scenarios. Within each scenario, situations occur where different decision design options could have ethical consequences. Making these decisions reliant on human judgement is both counter-productive to the aim of exploring very large number of scenarios in a timely manner and infeasible when considering the workload needed to involve humans in each of these choices. In this paper, we move human judgement outside the simulation decision cycle. Basically, the human will design the ethical metric space, leaving it to the simulated environment to explore the space. When the simulation completes its testing cycles, the testing environment will come back to the human commander with a few options to select from. The human commander will then exercise human-judgement to select the most appropriate course of action, which will then get executed accordingly. We assume that the problem of designing metrics that are sufficiently granular to assess the ethical implications of decisions is solved. Subsequently, the fundamental problem we look at in this paper is how to weight ethical decisions during the running of these simulations; that is, how to dynamically weight the ethical attributes when agents are faced with decision options with ethical implications during generative simulations. The multi-criteria decision making literature has started to look at nearby problems, where the concept of entropy has been used to determine the weights during aggregation. We draw from that literature different approaches to automatically calculate the weights for ethical attributes during simulation-based testing and evaluation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    P. K. Davis, S. C. Bankes, and M. Egner, Enhancing strategic planning with massive scenario generation: Theory and experiments . Rand Corporation, 2007, vol. 392

    work page 2007

  2. [2]

    H. A. Abbass, Computational red teaming . Springer, 2015

    work page 2015

  3. [3]

    J. Tang, G. Leu, and H. A. Abbass, Simulation and Computational Red Teaming for Problem Solving . John Wiley & Sons, 2019

    work page 2019

  4. [4]

    Tzeng and J.-J

    G.-H. Tzeng and J.-J. Huang, Multiple attribute decision making: methods and applications . Springer, Berlin Heidelberg, 1981

    work page 1981

  5. [5]

    Bouyssou, T

    D. Bouyssou, T. Marchant, M. Pirlot, A. Tsoukias, and P. Vincke, Evaluation and decision models with multiple criteria: Stepping stones for the analyst . Springer Science & Business Media, 2006, vol. 86

    work page 2006

  6. [6]

    The analytical hierarchical process (vol. 1),

    T. Saaty, “The analytical hierarchical process (vol. 1),” 1980

    work page 1980

  7. [7]

    Electre: Une m ´ethode pour guider le choix en pr ´esence de points de vue multiples,

    R. Benayoun, B. Roy, and B. Sussman, “Electre: Une m ´ethode pour guider le choix en pr ´esence de points de vue multiples,” Note de travail, vol. 49, pp. 2–120, 1966

    work page 1966

  8. [8]

    Elaboration dinstruments daide a la decision. methode promethee,

    J. Brans, R. Nadeau, M. Landry et al., “Elaboration dinstruments daide a la decision. methode promethee,” Laide a la Decision: Nature, In- struments et perspectives Davenir, Presses de Universie Laval, Quebec, Canada, pp. 183–214, 1982

    work page 1982

Show all 25 references
  1. [9]

    Four methods for assessing multi-attribute utilities: An experimental validation,

    G. W. Fischer, “Four methods for assessing multi-attribute utilities: An experimental validation,” Ph.D. dissertation, University of Michigan, 1972

    1972

  2. [10]

    A comprehensive review of multiple criteria decision-making (mcdm) methods: Advancements, applications, and future directions,

    S. K. Sahoo and S. Goswami, “A comprehensive review of multiple criteria decision-making (mcdm) methods: Advancements, applications, and future directions,” Decision Making Advances , vol. 1, pp. 25–48, 12 2023

    2023

  3. [11]

    A judgment aggregation method for fuzzy multi criteria decision making,

    A. Anniciello and E. Masciari, “A judgment aggregation method for fuzzy multi criteria decision making,” in 2023 31st Euromicro In- ternational Conference on Parallel, Distributed and Network-Based Processing (PDP). IEEE, 2023, pp. 283–290

    2023

  4. [12]

    Determining the level of agreement of expert estimates,

    V . Podvezko, “Determining the level of agreement of expert estimates,” International Journal of Management and Decision Making , vol. 8, no. 5-6, pp. 586–600, 2007

    2007

  5. [13]

    Bossel, S

    H. Bossel, S. Klaczko, N. M ¨uller, and R. L. Keeney, Multiattribute Utility Analysis—a Brief Survey . Springer, 1976

    1976

  6. [14]

    Multi-criteria decision making (mcdm) methods and concepts,

    H. Taherdoost and M. Madanchian, “Multi-criteria decision making (mcdm) methods and concepts,” Encyclopedia, vol. 3, no. 1, pp. 77– 87, 2023

    2023

  7. [15]

    Multi-attribute utility theory: models and assessment procedures,

    D. V on Winterfeldt and G. W. Fischer, “Multi-attribute utility theory: models and assessment procedures,” The University of Michigan, Ann Arbor, Michigan, Tech. Rep., 1973

    1973

  8. [16]

    Value modelling for multi- stakeholder and multi-objective optimisation in engineering design,

    E. Papageorgiou, M. H. Eres, and J. Scanlan, “Value modelling for multi- stakeholder and multi-objective optimisation in engineering design,” Journal of Engineering Design , vol. 27, no. 10, pp. 697–724, 2016

    2016

  9. [17]

    Explicit methods for attribute weighting in multi-attribute decision-making: a review study,

    J. Pena, G. N ´apoles, and Y . Salgueiro, “Explicit methods for attribute weighting in multi-attribute decision-making: a review study,” Artificial Intelligence Review, vol. 53, pp. 3127–3152, 2020

    2020

  10. [18]

    Weighting methods for multi-criteria decision making tech- nique,

    G. Odu, “Weighting methods for multi-criteria decision making tech- nique,” Journal of Applied Sciences and Environmental Management , vol. 23, no. 8, pp. 1449–1457, 2019

    2019

  11. [19]

    Determining attribute weights based on het- erogeneous discriminating power and solution reliability in evidential reasoning approach,

    Y . Liu and M. Xue, “Determining attribute weights based on het- erogeneous discriminating power and solution reliability in evidential reasoning approach,” International Journal of Fuzzy Systems , vol. 23, no. 7, pp. 2235–2251, 2021

    2021

  12. [20]

    Determining attribute weights for multiple attribute decision analysis with discriminating power in belief distributions,

    C. Fu, D.-L. Xu, and M. Xue, “Determining attribute weights for multiple attribute decision analysis with discriminating power in belief distributions,” Knowledge-Based Systems, vol. 143, pp. 127–141, 2018

    2018

  13. [21]

    Combination weighting method based on maximizing deviations and normalized constraint condition,

    F. Yin, L. Lu, J. Chai, and Y . Yang, “Combination weighting method based on maximizing deviations and normalized constraint condition,” International Journal of Security and Its Applications , vol. 10, no. 2, pp. 39–50, 2016

    2016

  14. [22]

    Effects of the entropy weight on topsis,

    P. Chen, “Effects of the entropy weight on topsis,” Expert Systems with Applications, vol. 168, p. 114186, 2021. 7

    2021

  15. [23]

    Decision support system perspective using entropy and multi-attribute utility theory in the selection of the best division head,

    M. W. Arshad, S. Sumanto, and S. Setiawansyah, “Decision support system perspective using entropy and multi-attribute utility theory in the selection of the best division head,” Jurnal Media Informatika Budidarma, vol. 8, no. 2, pp. 1109–1119, 2024

    2024

  16. [24]

    An integrated attribute-weighting method based on pca and entropy: Case of study marginalized areas in a city,

    O. Pliego-Mart ´ınez, A. Mart ´ınez-Rebollar, H. Estrada-Esquivel, and E. de la Cruz-Nicol ´as, “An integrated attribute-weighting method based on pca and entropy: Case of study marginalized areas in a city,” Applied Sciences, vol. 14, no. 5, p. 2016, 2024

    2016

  17. [25]

    C 4.5: Programs for machine learning,

    J. R. Quinlan, “C 4.5: Programs for machine learning,” The Morgan Kaufmann Series in Machine Learning , 1993. APPENDIX A STEP -BY-STEP CALCULATIONS FOR INFORMATION -T HEORETIC WEIGHT DERIVATIONS This appendix presents detailed calculations for deriving various information-theo...

    1993

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.