Programme decision sequences · Essai 14
Learning how quickly the world changes
Two programme dashboards both show 20% risk today and a 50% chance of fast change. One programme should buy a report. The other should wait. Which important fact has the dashboard left out?
The previous essai chose a reporting cadence when the rate of change was known and every report was immediate and correct. Here the rate itself is unknown, reports can be wrong, and a report can arrive one window late. With the independent opening prior, either first signal leaves the fast-law probability at 50%. Later signals can reveal which conditions tend to coincide with fast change.
Have we stored enough to act?
Five shuttle windows remain. Protecting costs 3 units, an adverse window costs 10, and a report costs 0.35. The report is 90% accurate and arrives one window later. Both starting states have the same dashboard totals. Change only the association between today's condition and the persistent rate of change.
When trouble is disproportionately in slow-changing programmes, it is more likely to persist, while fast-changing programmes that are normal today can deteriorate. Waiting and reporting therefore lead to different expected costs in these two starting states. Neither joint mix can be reconstructed from its two headline percentages.
A festival shuttle, five decisions left
In each remaining window the shuttle either operates, risking ten loss units in an adverse condition, or protects the service at a cost of three. A report costs 0.35. Conditions may switch at a slow rate of 5% or a fast rate of 45%; one of these laws persists throughout this run. These are invented decision units, not calibrated financial estimates.
Try Same risk, slow law, then Same risk, fast law. Both begin with a 20% chance of an adverse condition. The slow-law programme buys a report; the fast-law programme does not. A delayed report about rapidly changing conditions may arrive too stale to justify its price.
Walk one observation history
The policy chooses whether to buy a report. If it buys, select an observed signal; its probability is shown. The hidden true condition is never revealed to the policy. When reports are delayed, the signal concerns the previous condition, and the belief has already been advanced to the next window.
Starting from the opening example below the dashboard test, choose Normal and then Normal. The fast-law belief falls to 39.02% and the policy buys another report. Restart and choose Normal then Adverse. It rises to 74.35% and the policy waits. Both histories start from the same independent prior.
What the extra state buys
The adaptive policy can condition later reporting decisions on earlier signals. Its advantage over a fixed calendar prices the freedom to change reporting dates. A separate comparison is needed to isolate the value of learning the rate itself. The best fixed calendar chooses all report dates before the run, but still uses every report correctly when choosing whether to operate. This is a demanding comparator: it shares the same uncertainty model and pays the same prices.
The frozen-law comparator replaces the two persistent possibilities with their average switching rate, then optimises its entire reporting policy. We evaluate its choices under the true two-law model. Its disadvantage measures a particular misspecification, not a universal penalty for simpler forecasts. The free-law benchmark reveals which switching law is true before the run, while leaving current conditions uncertain.
Knowing only today's adverse probability loses the correlation between the hidden law and the hidden condition. Repeated noisy reports can identify a stable pattern or a volatile one. In general, the useful decision state is the joint belief over condition and law.
Construction, proof and limits
Klenske and Hennig's dual-control study motivates including uncertainty about dynamics in the decision state. Their continuous approximate methods are not implemented here. This is our small, fully enumerated belief-state decision problem.
Each law is constant. The condition flips independently each window conditional on that law. A binary report is correct with the selected probability, independently conditional on the sampled condition. Bayes' rule updates the joint belief. Backward recursion compares buying with waiting, including all subsequent contingent decisions. Enumeration is exact for this finite model up to floating-point tolerance.
Independent checks
A separate oracle enumerates complete hidden law/condition trajectories and conditions directly on report likelihoods. It does not call the production transition or observation functions. It agrees in 192 parameter cases, including uninformative/perfect reports, zero price and both delays; every fixed calendar is also checked for horizons up to three. A paid report arriving after the final decision is never bought. Free revelation of the law is a lower cost benchmark.
Strongest objection. The information action is separate from the operating action. Operating creates no observed feedback, and neither operating nor protecting changes the hidden transition law. This therefore studies adaptive sensing and parameter learning, not the full dual effect of productive actions. Unknown accuracy, drift in the law, irregular delays and model error could overturn a fitted recommendation. It establishes why a reporting state needs more than a risk score; it does not calibrate a real programme's cadence.