Hangman AI: Transformer-Based Game Solver
An AI agent for the Hangman word game that integrates a custom character-level transformer with heuristic decision layers to outperform frequency baselines (≈18%) and pure neural approaches.
- Model: 6-layer encoder transformer (512 hidden, multi-head attention) trained with masked character prediction to infer missing letters.
- Data Expansion: Morphological augmentation via WordNet expanding vocabulary from ~250k to 750k+ variants for broader generalization.
- Strategy Engine: MasterGuesser module combines neural likelihood scores, dynamic opening letter ordering, and endgame pruning logic.
- Filtering: On-the-fly candidate dictionary reduction using pattern constraints + positional frequency recalculation.
- Performance: Full pipeline (transformer + augmentation + heuristics) sustains >60% solve rate over 1000+ randomized games.
- Artifacts: Includes training notebooks, inference pipeline, pretrained weights, and evaluation scripts for reproducibility.
Tech: Python, PyTorch, NLTK, Transformers, NumPy.
Evaluation harness records per-turn entropy reduction and guess efficiency metrics for model iteration.