State, memory and the next question

Enough to
hand over?

Toy scenario

You inherit a festival shuttle decision. The handover tells you the day and the chance of high demand. Is that enough to carry on?

The earlier experiment asks whether to reserve one bus, reserve three, or first collect more evidence. This one changes the handover. We compare different routes to the same belief, then ask whether they still permit the same next moves.

The distinction is small but useful to inspect: “What can we do next?” and “How did we get here?” need not require the same information.

Finite model · service-value points · handover benefit remains untested.

01 / Put the handover to a test

Same day. Same belief.
Same possibilities?

Change one rule to see what the compact state leaves out.

Which decision rule?

Loading the computed comparison…

Reading the finite experiment.

03 / Ask a different question

How did we get here?

The day and current belief do not reveal which surveys were bought, which signals arrived, or what was already spent. A compact state can preserve the future decision while losing the route that produced it.

Adding remaining allowance lets us recover survey spending from the known initial allowance, but still loses the ordered evidence history. The experiment keeps that history separately: open “Reveal source history” above to inspect the two routes.

Check the result across the six existing settings

The six settings from the earlier experiment are retained. The check asks whether one feasible policy for every possible future signal can be optimal after every history sharing a proposed state. It includes routes the recommended policy would not choose.

Computed counts will appear when the results load.

Does every shared-state group admit a common feasible policy with zero loss of expected future value?
Existing settingOriginal: (t, p)Allowance: (t, p)Allowance: repaired state

“Pass” concerns the continuation contract in this finite model, not the recoverability of history or the quality of a real handover.

What was changed, computed and left outside

A single rule changes

The original shuttle model carries the review number t and current high-demand belief p. Known model settings remain fixed. Past fees and waiting costs do not constrain subsequent choices.

The extension caps total survey fees at twice the fee for one survey. Only survey fees consume that allowance. The payoffs, evidence assumptions, waiting costs and six existing parameter settings are retained.

A survey remains possible only if it can arrive by the deadline and the remaining allowance covers its fee. A zero-fee survey consumes no allowance; the review deadline still bounds the experiment.

A finite continuation check

The calculation enumerates feasible histories and solves the remaining decision by backward induction. For each shared-state group it finds the best contingent policy feasible after every member, then compares its expected future value with the best value available to each member separately.

The original pair is checked under the original rule. The same pair is then challenged under the allowance rule, and the check repeated with remaining allowance added.

The model uses exact rational arithmetic; this page rounds decimal displays. This is a bounded policy comparison for the listed finite settings, not a general search for the smallest state or a measured improvement in human decisions.

Source and construction

Warren B. Powell’s Sequential Decision Analytics and Modeling (2022), §1.4, supplies the framing of state, decision, new information, transition and objective. His separate paper On State Variables, Bandit Problems and POMDPs discusses the role of state variables. The shuttle example, survey allowance and handover comparison are constructions of this essai, not results or a case study attributed to Powell.

Three nearby questions

Carry the question a little further.

Knowing Enough asks how the available actions change which distinctions a decision needs.

Bricklaying a Rotunda sets a stronger preservation requirement for an open interface through which construction must continue.

Enough for the next question separates a shared verdict from the different reasons that may support it.