Decision Algorithms

Which decision rule fits the choice you actually face?

An algorithm is a constraint on judgment, not a substitute for it. Choose one only after naming the decision, time horizon, reversibility, uncertainty, and feedback speed.

Choose By Decision Shape

Situation	Useful rule	Use when	Watch for
Search with a deadline	37% rule	Options arrive in sequence and recall is costly	The assumptions rarely fit high-stakes or tiny samples
Learn while earning	Explore/exploit	Each trial returns useful feedback	Exploring forever or exploiting a lucky result
Repeat after success	Win-stay, lose-shift	Outcomes arrive quickly and conditions stay comparable	Noise masquerading as failure
Protect discovery	Upper Confidence Bound	Unknown options may be better than proven ones	Optimism without a loss limit
Sequence finite work	Weighted shortest processing time	Value and duration can be estimated	Urgent small work crowding out the critical path
Estimate from sparse observations	Laplace's rule of succession	You need a bounded prior from few trials	False precision and changing conditions
Coordinate strategic actors	Nash equilibrium	Each actor's best move depends on the others	Treating a stable outcome as a valuable one
Break false precision	Randomization	Options are close or exploration needs protection	Using chance to avoid accountable judgment

Practice

1. State the choice and classify the door: reversible, one-way, or compound.
2. Name the time budget, loss limit, unknowns, and feedback interval.
3. Choose one rule and write why its assumptions fit.
4. Let a named human accept, revise, or reject the recommendation.
5. Record the prediction and review date in the decision journal.

If one dependency controls time to value, identify the critical path before optimizing the rest. If the choice itself remains unclear, ask which uncertainty could change it.

Failure Modes

• Algorithm theater — applying a named rule without checking its assumptions.
• False precision — turning weak estimates into confident scores.
• Method shopping — choosing the rule that supports a preferred answer.
• No feedback — using an adaptive rule where outcomes cannot be observed.
• Authority loss — treating a recommendation as consent.

Checks

• The rule's assumptions match the decision shape.
• A named human accepts, revises, or rejects the recommendation.
• The journal freezes the prediction, falsifier, and review date.

Changes my mind: the same decision quality is achieved when the method's assumptions, authority, and feedback interval remain implicit.

Context

• depends-on Decision Making — define the decision before selecting its rule.
• applies-to Decision Graph — make the chosen rule, assumptions, authority, and result queryable.
• pairs-with Matrix Thinking — explore the option space before converging.
• contrasts-with Essential Algorithm — decision rules choose a route; the essential algorithm repeatedly routes system work.

Sources

• Algorithms to Live By summary
• Explore or Exploit

Questions

Next question: which assumption would make your chosen rule unsafe?

• How quickly will reality return feedback?
• Who can stop or revise the choice?
• What will you record so the next decision starts wiser?