Evidential Decision Theory¶
A decision rule that prefers the act with highest outcome utility conditional on that act occurring, even when this diverges from causal-effect ranking.
Core Idea¶
Evidential decision theory (EDT) recommends choosing the act whose occurrence would be the best evidence for a desirable outcome. Formally it compares acts using the expectation of outcome desirability under P(outcome | act). The conditional bar matters: the theory asks what one should expect to learn about the world upon learning that one performs a given act.
That is not the same as causal decision theory's question about what performing the act would bring about. The frozen Newcomb case makes the split visible because a reliable predictor filled a box before the choice: choosing one box is evidence for a pre-existing high payout, whereas taking both cannot now cause its contents to change. EDT's prescription follows from the stipulated correlations and utilities; whether such evidential ranking is rational is a live philosophical dispute, not a proved universal rule.
Structural Signature¶
Sig role-phrases:
- Available acts — The alternatives over which the agent must choose. It is constitutive. Counterfactual: Without a feasible act set there is no EDT recommendation.
- Conditional outcome model — P(O|A) says how likely an outcome is given news that an act occurs. It is constitutive. Counterfactual: Replacing it with causal intervention probabilities changes the theory.
- Outcome desirability — D(O) supplies values to probability-weight under each act. It is constitutive. Counterfactual: Conditional probabilities alone cannot rank actions without valued outcomes.
- Expected-utility comparison — Ranks acts by the sum of conditional outcome probabilities times their desirabilities. It is constitutive. Counterfactual: A mere correlation observation without maximizing the weighted value is not EDT's choice rule.
- Causal contrast — Marks when an act predicts an outcome without causing it, producing a possible EDT/CDT divergence. It is boundary. Counterfactual: The theories often agree when evidential and causal rankings coincide.
What It Is Not¶
- Causal decision theory. CDT substitutes a counterfactual causal effect distribution for P(outcome | act); the rankings can diverge.
- Prediction alone. An act can predict a result without a utility scale or an EDT maximization rule.
- Newcomb's paradox itself. The scenario tests EDT but is not the theory's definition or its sole possible application.
- Empirical finding. A formal prescription from stipulated probabilities is not evidence that people actually decide this way or should accept the norm.
- Closest near-miss. In Newcomb's setup taking only the opaque box predicts a large prefilled reward, although choosing now does not causally refill it; this separates evidential from causal evaluation.
Scope of Application¶
- Decision-theory comparison. Contrast conditional and causal probability models under the same acts and outcomes.
- Newcomb analysis. Explain the one-box recommendation from the predictor correlation and stipulated payoffs.
- Correlated-agent cases. Examine when one's action is evidence about another similar agent's action without direct influence.
- Normative critique. Locate objections to choosing on favorable news rather than causal production.
Clarity¶
Write the conditional P(O|A) and outcome desirability explicitly before claiming an EDT verdict. Do not replace conditional evidence with an intervention probability, and do not treat a stipulated predictor as an observed phenomenon.
Manages Complexity¶
The single conditional expectation compresses a complete act–outcome distribution into one ranking. That formal economy hides a philosophical choice: whether predictive correlation should guide action even when the chosen act does not cause the favorable correlated state.
Abstract Reasoning¶
- List the available acts and outcomes without changing the case after comparison begins.
- Assign each outcome a desirability and each act a conditional outcome distribution.
- Compute the sum of conditional probability times desirability for each act.
- Choose the highest conditional expected value as EDT's prescription.
- Compare with the separate causal counterfactual model and identify the source of divergence.
Knowledge Transfer¶
The conditional-outcome ranking transfers to other stipulated decision problems with acts, probabilities, and utilities, including correlated-agent puzzles. Newcomb's numerical recommendation does not transfer without its predictor/payoff assumptions, and the formal rule's application does not settle the normative EDT-versus-CDT dispute.
Examples¶
Canonical¶
Under the frozen 99%-accurate predictor and prefilled-box payoffs, observing that one boxes is evidence for the million-dollar content, so EDT ranks one-boxing above two-boxing. This is a stipulated thought experiment, not an empirical study of human choice.
Mapped back: Available acts → one opaque box versus both boxes; Conditional outcome model → one-box action is highly correlated with full opaque box; Outcome desirability → specified money payoffs; Expected-utility comparison → conditional expected value favors one box; Causal contrast → present action does not change already-filled box.
Applied / In Practice¶
The frozen article's published twin-prisoner's-dilemma thought experiment compares cooperation with defection when an agent takes her choice as evidence of a psychologically identical twin's choice. Under the stipulated perfect-correlation payoffs, EDT favors the mutually cooperative outcome; this is a real scholarly use of the decision rule, not field evidence that actual twins reason alike.
Mapped back: Available acts → cooperate or defect; Conditional outcome model → choice predicts the analogous twin's act under stipulation; Outcome desirability → individual payouts in the published case; Expected-utility comparison → cooperation ranked by conditional expectation; Causal contrast → evidential correlation is not a causal influence across rooms.
Structural Tensions¶
T1 — Evidential Correlation versus Causal Efficacy. An act can be excellent news about a pre-existing reward without producing that reward.
Diagnostic: Which probability model is doing the ranking in this case?
T2 — Formal Recommendation versus Contested Rationality. The theory produces precise verdicts in stipulated cases while critics dispute whether auspicious evidence is a reason to act.
Diagnostic: Is the stated result a deduction from EDT or an independent defense of its normativity?
Structural–Framed Character¶
The approved DAG parent is Expected Utility: acts are ranked by probability-weighted valued outcomes. EDT narrows the probability to evidential P(outcome|act), rather than a causal intervention distribution.
Evaluative weight: Normative ranking is explicit, but the EDT-versus-CDT dispute is not settled by the definition. Human-practice-bound: Moderate, because outcomes and utilities are specified by a decision maker while the rule is formal. Institutional origin: Decision theory formulated the criterion, not a single institution. Vocabulary travels: It applies to stipulated choice problems with act-conditional probabilities; Newcomb outcomes depend on case assumptions. Import versus recognize: Recognize EDT by its conditional-on-act calculation; substituting causal probabilities imports a different rule.
Its character: A formal decision-rule subtype with portable expectation logic and a precise evidential conditional.
Structural Core vs. Domain Accent¶
Skeletal core. Rank uncertain acts by probability-weighted outcome desirability.
Domain-bound accent. EDT uses P(outcome|act) as evidence about outcomes, not an intervention probability.
Why not prime. Expected utility is broader; without the evidential conditional this is another decision rule.
Instantiates / Related Primes¶
This entry is a kind of Expected Utility.
-
Strict parent — Expected Utility. EDT ranks risky acts by probability-weighted desirability; its differentia is conditioning on the occurrence of each act rather than on its causal intervention.
-
Neighbor — Causal decision theory. It uses a counterfactual effect model and may disagree even while evaluating the same choice set.
Relationships to Other Abstractions¶
Current abstraction Evidential Decision Theory Domain-specific
Parents (1) — more general patterns this builds on
-
Evidential Decision Theory is a kind of Expected Utility Prime
EDT ranks acts by probability-weighted outcome desirability, specializing expected utility with P(outcome | act).Live expected_utility ranks risky options by probability-weighted utility. EDT has exactly that act/outcome/utility/ranking structure, with the narrower rule that probabilities are evidential conditionals P(O|A), not causal intervention distributions. Expected utility can occur without EDT's conditional-on-act criterion; the child-to-parent subsumption is strict.
Hierarchy paths (4) — routes to 3 parentless roots
- Evidential Decision Theory → Expected Utility → Expected Value → Aggregation → Micro Macro Linkage
- Evidential Decision Theory → Expected Utility → Preference
- Evidential Decision Theory → Expected Utility → Expected Value → Probability → Measure → Set and Membership
- Evidential Decision Theory → Expected Utility → Expected Value → Probability → Measure → Aggregation → Micro Macro Linkage
Neighborhood in Abstraction Space¶
Evidential Decision Theory sits in a crowded region of the domain-specific corpus (35th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.
Family — Decisions Under Constraint & Commitment (9 abstractions)
Nearest neighbors
- Illusion of control — 0.90
- Two-Moment Decision Model — 0.89
- Buridan's ass — 0.88
- Non-Consequential Reasoning — 0.88
- Causalism — 0.87
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
- Causal decision theory. Tell: Is P(O|A) being replaced by the effect of doing A?
- Bayesian conditioning alone. Tell: Is an action actually selected by conditional expected utility?
- Newcomb's scenario. Tell: Is a test case mistaken for the general theory?
- Psychological prediction. Tell: Is a normative recommendation being misreported as observed human behavior?
References¶
- Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Evidential_decision_theory (revision 1363960459).
- Preserved source candidate: https://www.econstor.eu/bitstream/10419/220553/1/cmsems-dp0194.pdf
- Preserved source candidate: https://philpapers.org/rec/GREACW-2
- Preserved source candidate: https://philpapers.org/rec/CONADB
- Preserved source candidate: https://globalprioritiesinstitute.org/wp-content/uploads/2019/MacAskill_et_al_Evidentialist_Wager.pdf
- Preserved source candidate: http://www.informaworld.com/index/739194078.pdf
- Preserved source candidate: https://archive.org/details/foundationsofcau0000joyc/page/n7/mode/2up
- Preserved source candidate: http://plato.stanford.edu/entries/decision-causal/
The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.