Optimal Foraging TheoryA familiar patch and an untested one carry different returns, search costs and uncertainties.
What It Is
An animal feeding in one place eventually faces a decision: keep extracting the resources that remain, or spend energy searching somewhere else. Staying has a known return that may be declining. Leaving costs energy before there is any assurance of finding food.
Optimal foraging theory (OFT) models this resource-acquisition problem. Developed in behavioral ecology during the 1960s and 1970s, it examines what animals eat, where they search and when they move. The organism has finite energy, resources are unevenly distributed, and their locations are visible only through imperfect cues. A useful search must acquire resources before the reserves run out.
The objective is to maximize energy gained minus energy spent searching, without exceeding the available budget. The same shape of problem appears in startup survival: runway supplies the finite reserve, product-market fit is unevenly distributed across possible markets, search consumes runway and signals about fit are imperfect. Each is resource-constrained search under uncertainty. A forager's intake can be counted in calories, while what a startup learns in a month has no such unit.
A closer look
A choice under limited reserves
The article uses foraging to reason about exploration and exploitation; the useful choice depends on costs and changing returns.
Read this diagram
Inspect the current return branches into Keep using this patch; Spend resources searching.
The Central Trade-Off
Exploitation extracts resources from a known location. The gain is known because the resource has been discovered. The organism is already there, so the energy cost is low. The risk is depletion.
Exploration searches for a new location. Travel and search raise the cost, while the possible gain remains unknown. The organism may find nothing and exhaust its energy.
For a forager, the comparison can be written in units of energy:
E_total = E_current_patch × P_patch_remains - C_extraction
vs
E_total = E_new_patch × P_find_patch - C_search - C_travel
Where:
E = expected energy gain
P = probability
C = cost in energy units
Leaving is worthwhile when the expected return from a new search exceeds the return from continuing in the current patch. The marginal value theorem gives a rule for that transition.
The Marginal Value Theorem
An organism should leave a patch when its marginal rate of energy gain falls below the average rate available across the habitat. The relevant quantity is the return on the next interval of effort, rather than the total already collected.
Energy ↑
Gain | ╱───── (Diminishing returns in patch)
| ╱
| ╱
|╱___________________
0 Time spent in patch →
Leave when: (dE/dt)_patch < (E/T)_average_habitat
For a startup, the corresponding comparison is information gained per month of runway. Continue in the current direction while the next month there is expected to teach more than the average month after a pivot. When local insights diminish and another direction is expected to teach more, pivot.
The rate alone does not settle the decision. Confidence and the cost of validation also affect the expected return. Suppose a founder compares the two directions and finds this:
| Metric | Current direction | New direction | Implication |
|---|---|---|---|
| What a month teaches | Little that is new | Much more, if the direction is real | Favors pivoting |
| Confidence in direction | Fairly high | Low | Favors staying because uncertainty is costly |
| Runway required to validate | About 2 months | About 4 months | Favors staying because validation costs less |
An expected-value judgment combines these factors under the remaining resource constraint.
Staying 12 minutes keeps working a patch that now yields more slowly than moving on would. That earns 85% of the best overall rate, which comes from leaving at about 6 minutes.
Patch Choice and Search Patterns
A search can use no cues, environmental cues or other searchers' results. Each changes how much space must be examined.
| Strategy | Search behavior | Cost | Efficiency |
|---|---|---|---|
| Random search | Explore space uniformly | Grows with the whole area | Low: empty and resource-rich areas receive equal attention |
| Informed search | Concentrate where cues predict resources | Grows only with the area the cues point to | High when cues correlate with resources |
| Social information | Observe where others found resources | Watching, then traveling | Very high: search begins from confirmed resources |
The mosquito model combines heat, CO2 and movement to predict where blood is available. That is informed search. Five mosquitoes already feeding at a location provide social information: blood has been found there, rather than merely predicted from a cue.
Startups have corresponding search strategies:
| Strategy | Example | Information/cost ratio |
|---|---|---|
| Random | Build random features and hope something works | Very low |
| Informed | Build from user research and behavioral data | Medium |
| Social | Copy a validated model with a small differentiation | High |
| Hybrid | Validate similarly to competitors, then differentiate where sensors reveal opportunity | Very high |
The hybrid approach combines the social and the informed strategies. Others' product-market fit narrows the search space. Further observation can then locate an underserved niche inside that validated area. Following a demonstrated gradient and testing cheaply before committing resources preserves more of the search budget than beginning randomly.
Sensor Reliability and Search Efficiency
Search efficiency depends on how accurately a cue predicts a resource. Two rates describe a cue. The true-positive rate is how often the cue is present when the resource is there. The false-positive rate is how often the cue is present when the resource is not.
A larger difference between the true-positive and false-positive rates lowers expected search cost, because fewer visits are wasted. When the two rates are equal, the cue tells the searcher nothing. Poor calibration sends the organism to locations where the resource is absent, consuming energy without a return.
For startup sensors, the comparison is:
| Sensor | Signal despite no PMF | Calibration quality |
|---|---|---|
| Behavioral usage data | Seldom | High |
| Payment behavior | Rarely | Very high |
| Word-of-mouth growth | Sometimes | High |
| Expressed interest | Often | Very poor |
| Hypothetical commitment | Most of the time | Extremely poor |
Expressed interest and hypothetical commitment appear readily where there is no fit. A search steered by those two is drawn toward locations where the resource is absent, which is systematic misdirection. Integrating multiple sensors reduces false positives and makes the search better match the resource distribution.
The Giving-Up Time Problem
Leaving too early discards information still available in the patch. Staying too long spends the remaining reserve on a depleted location. The optimal giving-up time balances the opportunity cost of staying against travel cost and the risk that the next patch is worse. The cost of staying is what better patches would have yielded in the same time, and it grows the longer the stay. The cost of leaving is the travel plus the chance that the next patch turns out worse. The time to give up is when the first has grown to match the second.
For a startup, the persistence rule has the same form. Persist while the current direction is expected to teach more over the remaining runway than a new direction would over what is left of that runway after the switch. The switch has a cost of its own, paid in runway: the context switch and a new learning curve.
Four judgments are required. Is the current direction still producing insights? What learning rate is realistically available in another direction? How much runway will the switch consume before reaching an equivalent information state? How much runway remains to carry out either strategy?
A premature pivot abandons a useful gradient before its local information has been extracted. A late pivot consumes runway after that information is exhausted. Timing depends on an honest estimate of all four variables.
Integration with Mechanistic Framework
A forager's decision is an expected value comparison under a resource constraint:
Execute if and .
The expected-value account of motivation weighs the same kinds of input: the size of a reward, its probability, and the effort and delay of getting it. Both allocate resources under uncertainty.
Prevention architecture applies the same approach to attention. Removing frequent, low-value temptations such as phones and notifications excludes them from the search space. This resembles habitat selection: avoiding predator-rich or resource-poor places confines the search to areas with a higher yield.
Related Concepts
- Startup as a Bug applies OFT to organizational survival.
- Cybernetics describes control and feedback under constraints.
- Expected Value provides the resource-allocation calculation.
- Information Theory measures information's contribution to reducing uncertainty.
- Prevention Architecture constrains search through habitat selection.
- Activation Energy describes the cost of starting something, which a change of direction also carries.
- Question Theory examines how a question bounds a search.