Finite additive MCDA information value

This experimental workflow values perfect information inside a finite, compensatory additive multi-criteria decision model. The request fixes the alternatives, criterion directions and raw units, ex-ante value-function anchors, normalized nonnegative weights, and the joint outcome/preference law before any information is resolved.

For alternative a and finite state omega, the additive value is

U(a, omega) = sum_k w_k(omega) v_k(x[a, k, omega])

The current value optimizes expected utility before learning a partition. Each information action then conditions on its declared criterion, preference, or joint partition and re-optimizes. Gross information value is the resolved value minus the current value; net value additionally subtracts that action’s declared, disjoint cost.

Run the exact correlated-state fixture

From a source checkout, run:

voiage --format json calculate-mcda-information \
  specs/frontier/mcda-information/v1/fixtures/normative/input.json \
  --output mcda-information-result.json

The four-state synthetic fixture preserves correlation between criterion performance and preferences. Criterion-only and preference-only resolution do not change the baseline choice, so each has zero gross value. Joint resolution changes the choice in one state and has gross value 0.028, entirely recorded as an interaction in the declared normalized decision-value unit. Costs remain visible, so a zero gross value may correctly have a negative signed net value.

The result also preserves complete choice ties, full rankings, statewise and expected regret, fractional complete-tie rank acceptability, and expected and statewise Pareto diagnostics. It never silently renormalizes probabilities or weights and never moves the fixed value-function anchors after information is observed.

Python and accessible plots

import json
from pathlib import Path

from voiage import mcda_information_value
from voiage.plot import (
    plot_mcda_information_value,
    plot_mcda_rank_acceptability,
)

request = json.loads(
    Path(
        "specs/frontier/mcda-information/v1/fixtures/normative/input.json"
    ).read_text()
)
result = mcda_information_value(request)

value_ax = plot_mcda_information_value(result)
value_ax.figure.savefig("mcda-information-value.png", bbox_inches="tight")

rank_ax = plot_mcda_rank_acceptability(result)
rank_ax.figure.savefig("mcda-rank-acceptability.png", bbox_inches="tight")

The information-value chart distinguishes criterion, preference, and joint resolution with hatching, outlines, labels, and numeric annotations rather than colour alone. The rank-acceptability chart uses distinct markers and a labelled legend and explicitly identifies probabilities as fractional-tie probabilities.

Version 1 deliberately excludes AHP elicitation, outranking and veto methods, non-compensatory or nonlinear aggregation, post-information normalization, ordinal labels treated as cardinal weights, imperfect/sample information, social-choice aggregation, and endogenous feasible sets. Python is the only executable surface. Scientific review, stable promotion, release, and Rust/R/Julia parity remain separate gates; Mojo remains external.