End-to-End Implicit Learning of Logical Constraints from Raw Pixels via CNNs in Shidoku Shidoku Çözümünde Ham Piksellerden Mantiksal Kisitlara Uçtan Uca ESA ile Örtük Ö?grenme
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636742
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: Constraint Satisfaction Problems, Convolutional Neural Networks, Deep Learning, Logical Gap, Neuro-Symbolic AI, Shidoku
- Maltepe Üniversitesi Adresli: Evet
Özet
This study investigates solving Constraint Satisfaction Problems via end-to-end deep learning using 4 × 4 Shidoku puzzles. While existing approaches use computer vision to recognize inputs and apply rule-based algorithms to solve them, we trained a pure Convolutional Neural Network (CNN) to generate solutions directly from pixels without any explicit logical rules. Trained on 80% of a 1,351,300 synthetic image dataset, the model achieved 97.96% perfect puzzle accuracy on the 20% holdout test set (270,260 samples). Analysis shows performance heavily depends on clue density; the perfect solve rate of 99.17% on 6-clue puzzles drops to 86.64% on 4-clue puzzles (the theoretical minimum). Error distributions empirically prove the existence of a "logical gap,"demonstrating that statistical models struggle to satisfy hard constraints as information entropy increases.