STAR Summary
- Situation
- I wanted to see if an RL agent could learn to solve Nonogram puzzles.
- Task
- Owned the end-to-end build, from implementation through the final deliverable.
- Action
-
- Generated large batches of 5x5 puzzles and trained a hybrid CNN + Transformer policy network to solve them.
- Shaped rewards around unique guesses, row/column completions, and full-board solves to guide exploration.
- Result
-
- Evaluated on generated 5x5 boards, but the published materials do not yet include a reproducible held-out solve-rate result.
- Stack
- Python · PyTorch · AWS · Docker
- Status
- Live interactive demo
Evaluation & tradeoffs
- Status
- Not benchmarked
- Goal
- Measure whether the learned policy can complete unseen generated 5x5 Nonogram boards.
- Dataset
- Generated 5x5 Nonogram puzzles.
- Split
- The project references held-out boards, but split sizes, seeds, and a reproducible evaluation set are not published.
- Baseline
- No rule-based, search-based, or random-policy baseline is published.
- Decision
- Retained the hybrid CNN and Transformer policy for the live step-by-step solver demonstration.
- Limitations
-
- A held-out solve rate with board count, seed, and confidence interval has not been published.
- Results on generated 5x5 boards do not establish performance on larger or human-authored puzzles.
- Evidence
- Project repository
Demo
Generate a new Nonogram puzzle, then watch the agent solve it step-by-step.
- Click “New Puzzle” to generate a board.
- Once it’s ready, the button switches to “Solve With AI”. Click it to start solving.
- Use “Solution” to reveal the full answer key if you want to compare.
- If AWS is warming up, wait for the status to turn ready and try again.
Links
Notes
The solver focuses on generated 5x5 puzzles; scaling to larger puzzles would need curriculum learning or search.
