ashutosh1919/neuro-symbolic-sudoku-solver

Implements Neural Logic Machines (NLM), a neural-symbolic architecture that grounds first-order logic predicates (isRow, isColumn, isSubMat) as learnable tensors to encode Sudoku constraints. Uses policy gradient reinforcement learning with Jacinle framework on PyTorch to train an agent that iteratively fills the grid by applying neural operators over premise tensors. Includes pre-trained models, integration with Kaggle's 1M Sudoku dataset, and evaluation scripts demonstrating high solve rates across difficulty levels.

No commits in the last 6 months.