C2 · Publication Volume 12

Exploration as Uncertainty Reduction

risk, value of information and staged investment

Learning goals

After this lesson, you should be able to identify the uncertainty that controls a decision, distinguish uncertainty reduction from data accumulation, calculate a simple expected-utility and information-value example, and design a staged test whose result can change an action. You should also be able to state why uncertainty cannot be reduced to one percentage when geological model, measurement, access and consequence uncertainties differ.

Exploration begins before a target exists. A search-space decision may ask whether a region is geologically permissive, whether a footprint is detectable under cover or whether two process models predict different observations. The useful question is not “What data can we acquire?” but “Which uncertain proposition currently separates the available actions?”

Types of uncertainty and ignorance

Aleatory variability describes variation treated as inherent under the chosen model, such as local heterogeneity within a domain. Epistemic uncertainty reflects incomplete knowledge: an unmapped contact, unknown cover thickness, uncertain source fertility or ambiguous anomaly origin. Model uncertainty concerns the set and structure of explanations themselves. Decision uncertainty includes the consequences of acting under those states. The categories overlap, but they suggest different responses.

More samples can reduce uncertainty about a mean only if the sampling model, support and population are appropriate. More samples from the wrong medium do not resolve a transport process. Denser measurements do not fix a coordinate shift. A precise inversion does not establish that the selected physical property is diagnostic. Unknown unknowns cannot be represented honestly by adding a narrow error bar to a preferred model.

Maintain an uncertainty register with: proposition, current alternatives, evidence, consequence, reducibility, proposed test and review trigger. Do not force every entry into a probability. Ordinal confidence, scenario bounds and explicit ignorance can be more truthful than a numerical value derived from weak calibration.

Staged decisions and option preservation

Staging separates reversible, low-disturbance learning from irreversible or high-consequence action. Early work may compile existing data, verify locations, map exposure and test sample media. Later work may increase resolution or physical disturbance only when the result could justify it. Each stage should preserve useful options while retiring hypotheses that conflict with adequate observations.

A stage is defined by a decision, not by a traditional technique name. The same mapping method can be regional reconnaissance, detailed target testing or post-test reinterpretation. A drill hole can test stratigraphy rather than mineralisation. Conversely, a carefully designed surface survey can be a decisive target test when competing hypotheses predict opposite spatial patterns.

Stopping is a legitimate learning outcome. A program should stop or change when a necessary condition is adequately tested and absent, when evidence quality cannot support discrimination, when constraints make the proposed work unsafe or unlawful, or when another test dominates in expected information. Continuing because resources have already been spent is a sunk-cost error.

Information value and action change

Uncertainty narrows through tests only when observations can change an action
Uncertainty narrows through tests only when observations can change an action

Information has decision value when different possible outcomes lead to different preferred actions. If every result would lead to the same action, the test may improve description but has no immediate decision value. This does not make the science worthless; it means the decision objective should be stated separately.

For actions a, states s and utility U(a,s), the best current expected utility is

$EU_0=\max_a\sum_s P(s)U(a,s).$

With perfect information, the state is known before acting:

$EU_{PI}=\sum_sP(s)\max_aU(a,s),\qquad EVPI=EU_{PI}-EU_0.$

An imperfect test has possible results r. Its expected value uses posterior state probabilities and subtracts every relevant acquisition burden. No test can have net information value greater than perfect information under the same model. If a calculation violates that bound, the probabilities, utilities or accounting are inconsistent.

Designing a discriminating test

A discriminating test begins with predictions under alternatives. For each possible result, write what would be expected under each hypothesis, how likely the method is to detect it, and which action would follow. Include a barren or artefact explanation; otherwise every anomaly can be absorbed into a mineralising narrative.

Detection power depends on scale, orientation, support, noise and preservation. “No anomaly” has different meanings if line spacing is wider than the footprint, if transported cover disperses the signal, if the analytical method excludes the mineral host or if the survey was interrupted. Predeclare the adequacy condition for interpreting absence.

Prefer a test whose outcomes are separated across hypotheses and robust to plausible nuisance processes. A highly precise measurement of a feature predicted by all hypotheses has little discriminatory power. A modestly precise observation that one hypothesis requires and another forbids may be more valuable.

Worked synthetic example

A fictional target has two states: a coherent mineral system H with prior probability 0.35 and no coherent system N with probability 0.65. Two actions are available. Advancing without further information has utility +8 if H is true and −3 if N is true; stopping has utility 0 in either state. Utilities are unitless teaching values, not money.

The expected utility of advancing now is

$EU(advance)=0.35(8)+0.65(-3)=0.85,$

so advancing narrowly dominates stopping. If the true state were known, one would advance under H and stop under N, giving EU_{PI}=0.35(8)=2.80. The perfect-information bound is therefore EVPI=2.80-0.85=1.95 units.

Consider a synthetic test with P(+\mid H)=0.75, P(+\mid N)=0.20 and burden 0.35 units. The positive-result probability is 0.35(0.75)+0.65(0.20)=0.3925. The posterior after a positive result is 0.2625/0.3925=0.669; after a negative result it is 0.0875/0.6075=0.144. Advancing is preferred only after a positive result. Before test burden, expected utility is

$0.2625(8)+0.1300(-3)=1.71.$

After burden it is 1.36, an improvement of 0.51 over acting now and below the 1.95 perfect-information bound. The conclusion depends entirely on the synthetic prior, test performance and utilities. Its purpose is to show when a test changes action.

Interpretation workflow

  1. Name the immediate decision and the available actions.
  2. Separate geological state, observation process and consequence uncertainty.
  3. List at least two viable hypotheses plus an artefact or barren alternative.
  4. State the prior basis or use bounded scenarios if calibration is weak.
  5. Derive observable predictions at the method's scale and support.
  6. Specify what positive, negative and indeterminate results mean.
  7. Estimate whether each result would change the preferred action.
  8. Include acquisition burden, disturbance, safety and legal constraints.
  9. Compare the proposed test with a cheaper, safer or more discriminating alternative.
  10. Record the gate rule before viewing the result.

Practice and review

  1. Recalculate the example if the advance loss under N is −5. Does the preferred action before testing change?
  2. Design a test that is 90% sensitive but has a 45% false-positive rate. Explain why sensitivity alone is an inadequate quality measure.
  3. Give one geological, one measurement and one model uncertainty for a covered target.
  4. Describe a dataset that increases precision but cannot change the gate decision.
  5. Write an explicit adequacy condition under which a negative survey result would refute a pathway hypothesis.

Sources and further reading