We designed and implemented the first system for recognising similar
tactical problems in Crazyhouse, a chess variant in which dropping captured
pieces creates mating patterns that do not occur in classical chess. Each position,
together with its solution, is encoded as a document of tokens. From
roughly 17 million games we built a collection of 555,576 tactical problems,
all verified forced mates in three moves. Candidates are retrieved by BM25
and ordered by a trained re-ranker. In a web application, the domain expert
enters a tactical problem, the application returns candidates and the expert
labels them as similar or different. The data were collected in two phases
with distinct purposes: the first expert supplied the labels, the second the
features, derived from his own justifications following argument-based machine
learning. The contribution of these features is therefore measured on
the first expert’s labels, which the second never saw. Across 790 pairs, the
AUC rises from 0.733 to 0.847. An ablation shows that similarity is determined
by the dynamic context of the solution (AUC 0.789) rather than by
the static arrangement of the pieces (0.534). The limit of the system lies not
in ranking but in retrieval.
|
