Why test with puzzles?
A controlled puzzle can reveal a mechanism more clearly than an impressive application with many hidden assumptions.
The research needed problems with multiple valid or high-quality solutions, because the role of negative learning was not limited to finding one optimum. It also concerned whether the algorithm could discover more distinct solutions and preserve alternatives that did not resemble known poor structures.
The test suite therefore includes magic squares, combination chess puzzles and knight’s tours. These problems expose different patterns of shared, conflicting and deceptive substructures.
Magic squares: shared structure and conflicting building blocks
Magic-square experiments compare the edge-histogram baseline with positive COIN, negative COIN and the hybrid algorithm. The generator snapshots are especially informative: they show not only the final score but how the learned transition model differentiates relationships over time.
The deeper question is whether rewarding frequent edges is sufficient. If multiple valid solutions require conflicting structures, positive learning can concentrate too quickly. Negative learning may reopen alternatives by weakening relationships repeatedly associated with poor candidates.
Chess-placement puzzles: one board, different representations
The dissertation examines permutation and combination formulations across queens, rooks, bishops and knights. This is more than a collection of games. It tests how representation changes what constitutes a building block and how the model interprets conflict.
When the representation changes, the same objective may become a different learning problem. Algorithm comparisons without representation analysis are often comparisons of hidden assumptions.
Knight’s tour: long-range consequences from local decisions
A knight’s tour requires a sequence of legal moves covering the board. Each edge is locally meaningful, but early choices constrain distant parts of the tour. This makes the task useful for testing an edge-based generator and exposing the limits of first-order learning.
The chapter compares COIN and EHBSA performance, examines solutions generated by COIN, and studies generator snapshots. It also compares positive and negative variants to understand where each learning signal contributes.
What the dissertation claims
The dissertation concludes that negative correlation learning contributes to both the quality and quantity of solutions, but that its behavior changes with the amount of shared and conflicting building-block structure in each benchmark.
That conclusion should be read as evidence from the reported experimental suite—not as a universal theorem. The value of the work lies in connecting algorithm behavior to problem structure and in separating positive, negative and hybrid learning.
What students can extend
- Reproduce the original test suite in a current language and library stack.
- Measure diversity explicitly across distinct optima.
- Test adaptive reward and punishment coefficients.
- Investigate higher-order dependencies and graph-neural representations.
- Study when negative evidence removes deceptive structure and when it removes rare useful structure.