Find your next learning step
The next audit with the largest immediate value is not always the best first move. One scene may unlock a useful second audit, reveal a new stratum or calibrate a reference method used later. Non-myopic planning compares short sequences while keeping the eventual deployment decision visible.
Open the chapter contents · Remote Sensing root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the decision, or read from Chapter 1. The chapters remain open; these grouped routes are a navigable index.
Route 1 · Define the decision
Route 2 · Work the ledgers
Route 3 · Transfer the method
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE DECISION
1. One-step value can miss enabling evidence
A one-step knowledge gradient asks how much the next label is expected to improve the terminal decision. It can undervalue a calibration audit whose main benefit is making later scene labels more informative.
A lookahead policy evaluates paths: choose an audit, update beliefs, choose again and then act. Exact dynamic programming is rarely practical, so rollout and batch-informed approximations are used.
CHAPTER 2 OF 12 · DEFINE THE DECISION
2. A rollout needs a base policy
A rollout simulates possible first outcomes and then follows a simpler base policy for the remaining horizon. Its estimated value depends on the predictive model, horizon, base policy and terminal loss.
Use common random numbers or matched simulations when comparing first actions. Otherwise Monte Carlo noise can exceed the value difference between audit candidates.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. A one-step versus lookahead ledger
A team has two audits left. Cloud-edge calibration has modest immediate value but improves every later cloud-edge label.
| First audit | One-step net value | Expected best second audit | Two-step rollout value |
|---|---|---|---|
| Ordinary scene | 0.010 | Mountain shadow | 0.019 |
| Cloud-edge calibration | 0.006 | Cloud-edge validation | 0.028 |
| Mountain shadow | 0.014 | Ordinary scene | 0.021 |
The difference is an option value: calibration creates a better second-stage measurement path.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. Worked outcome-tree calculation
Suppose calibration passes with probability 0.7. A pass enables a validation audit worth 0.034; a fail redirects to mountain shadow worth 0.018. Calibration costs 0.004.
| Branch | Probability | Second-stage value | Weighted value |
|---|---|---|---|
| Pass | 0.70 | 0.034 | 0.0238 |
| Fail | 0.30 | 0.018 | 0.0054 |
| Total before first cost | 1.00 | — | 0.0292 |
| Net rollout value | — | 0.0292 − 0.004 | 0.0252 |
Recompute across plausible pass probabilities and downstream values. A rollout is only as credible as the simulated update paths.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Use batch plans as long-horizon guides
Batch design can approximate what a good long-horizon sequential policy wants to cover. A batch-informed first action selects one member of a diverse high-value set, then replans after its result.
This avoids committing the whole batch before outcomes arrive. It also exposes complementarity: two similar ordinary scenes may have less joint value than calibration plus a hard validation scene.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Stop when future paths no longer pay
Compare the best lookahead value with audit cost, delay and operational disruption. Stop when every credible path has low net value and the deployment action is stable across retained models.
Keep a reopen rule for sensor changes, new seasons or evidence that invalidates the predictive audit model. Non-myopic planning does not make old evidence permanent.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Changed-context transfer: coastal habitat maps
Transfer the method to mangrove classification where tidal calibration enables later shoreline labels. A one-step policy may repeatedly select clear inland scenes because their labels are easier.
The transfer check models acquisition windows, tide state, spatial clusters and reference-method error. A future scene that cannot be collected in the same tidal window should not appear as a guaranteed branch.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Connect the audit branch
The sequential-audit guide defines the one-step robust knowledge gradient and stability stop. The robust-value guide tracks uncertainty in the underlying action value.
Use the dependence-region guide when future audit value is driven by joint failure assumptions.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
A deeper lookahead is not automatically better. Model error and computation compound along simulated branches, while distant values may be smaller than their estimation noise.
Rollout improvement is relative to its base policy and simulated horizon. It is not a universal guarantee against all adaptive strategies.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Archive the horizon, candidate audits, base policy, predictive outcomes, update rule, terminal loss, costs, simulation seeds, variance reduction and stop criteria.
Report one-step and rollout rankings, Monte Carlo uncertainty, sensitivity to horizon and branch probabilities, and which first actions owe their value to enabling later measurements.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
How far ahead should I simulate?
Only as far as the model, budget and decision can support. Two or three steps often reveal complementarity without pretending to know a long future.
Can batch design replace rollout?
It can guide a first action efficiently, but a sequential policy should replan after each realised result.
Why might the easiest scene be low value?
Its label may confirm what is already known and leave the deployment boundary unchanged.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Frazier, Powell and Dayanik introduce the knowledge-gradient policy. Jiang and colleagues develop BINOCULARS, a batch-informed framework for efficient non-myopic sequential experimental design. Krause and colleagues study information-gain sensor placement with near-optimal guarantees under stated assumptions. The audit tree above is an original remote-sensing teaching example.
Continue through the Remote Sensing root, the Gold Standard of Education and How X Works.
Top · Contents · Previous · How X Works