Failure can reduce uncertainty
A failed solution is not automatically useful. Its value depends on whether we can identify the structure that contributed to the failure.
The dissertation draws on a constructivist view of knowledge: knowledge is not a perfect copy of reality, but a map of actions and relations that have proved viable for a goal. Negative knowledge extends that map with adverse routes, wrong turns and configurations that appear counterproductive.
This distinction matters. “Negative” describes the role of the information, not its moral value. A warning can be more valuable than another example of success when it prevents a large region of unproductive search.
Competence is partly the ability to act well. Expertise is also the ability to recognize the mistake before spending resources to perform it.
The Negative Building Block Hypothesis
The dissertation proposes the Negative Building Block Hypothesis (NBBH): new performance may be sought by avoiding short, low-order, low-performance schemas—negative building blocks. It is intended as a counterpart to the traditional Building Block Hypothesis.
In COIN, those structures are expressed through edge relationships in permutations. When an incidence repeatedly appears in below-average candidates, its probability can be reduced. When it appears in above-average candidates, it can be rewarded.
Why below-average solutions are not simply discarded
Most population methods preserve information from selected candidates and abandon the rest. COIN asks whether the rejected population contains a second signal. The objective is not to imitate failure, but to identify recurring structural evidence that can reshape the next sampling distribution.
Positive, negative and still unknown
Negative knowledge is not the exact complement of positive knowledge. The search space contains at least three regions:
- Known positive: structures observed in stronger candidates.
- Known negative: structures associated with weaker candidates.
- Unknown: combinations not yet supported by enough evidence.
Average candidates may contain both useful and harmful structures. Treating them as purely good or bad can prejudice the model. The selection mechanism must therefore discriminate between populations without pretending that every component inherits the label of the complete solution.
Four roles proposed for negative knowledge
- Push the search away from regions marked by repeated adverse structure.
- Support diversity while remaining dissimilar to candidates considered poor.
- Work with positive evidence to discriminate better and worse substructures.
- Help a constructive algorithm recognize and compose stronger solutions.
Where the idea becomes difficult
A low-fitness candidate can contain good parts. A strong candidate can survive despite a weak part. Correlation is not causation, and finite populations exaggerate accidents. Punishment that is too strong may erase useful diversity; punishment that is too weak may merely add computational ceremony.