Two small models. Real experiments. A place to watch what worked, understand what failed, and learn how game-playing AI is built.
A small neural policy learns expert shortcuts, then receives reinforcement-learning and search refinement. The goal: fill every playable cell, while using fewer movements.
Explore the Snake model120 / 120 full-board wins
Measured local starts across three 18 × 18 courses. On 60 fresh starts, it used 40.06% fewer movements than the cycle baseline.
Available here: verified recordings with pause, step, speed and seeking. Live neural inference is pending.
A spatial policy/value network learns from self-play and stronger search teachers. Combining the network with a proof controller produces our strongest local player.
Explore the Connect Four model87.5% against our Expert
The selected network + native search controller, measured on 100 games with varied starts and swapped seats. Draws receive half credit.
Research result. The neural network alone is much weaker. Browser ML play has not been qualified.
Each note follows a question, an experiment and a decision. Failed ideas stay visible because they teach us what to try—and when to stop.
Three experiments failed to establish a useful gain. We kept the champion and moved effort into a usable learning product.
Learning from deeper search improved some scores, but the Hard result was too uncertain to promote.
A 15.9% improvement in value error did not survive the playing-strength comparison.
Separate the neural policy, tree search and exact proofs before interpreting a win rate.
The selected Snake policy used 40.06% fewer moves than the cycle baseline on fresh measured starts.
Collecting twelve apples was not completing Snake. The real goal is to occupy every playable cell.