- Rules of the game
- History moves
- Color CLI
- R/W FEN (Forsyth–Edwards Notation)
- R/W PGN (Portable Game Notation)
- R/W SAN (Standard Algebraic Notation)
- Internal Engine (Iterative Deep Search)
- Engine support via UCI protocol
- Color GUI (Ebiten Library):
-
- Play Scene
-
- Editor Scene
-
- Analyzer Scene
-
- Settings Scene
- WASM Test-Build (it works LOL)
- My AI Engine (not integrated)
- Chess Model Intagration
I implemented a small internal chess engine that performs exhaustive depth-first move search and relies largely on a material-based evaluation. I trained AI model (as far as my resources allowed) to predict moves from FEN -> Move snapshots extracted from a filtered subset of Lichess games - sep 2025 dataset. Training ran for 46 epochs and lasted about 10 hours until EarlyStopping triggered (overfitting trigger).
Unfortunately the model was not integrated into the app, but you can download and test it locally from the repository directory. The model I trained can be found in the repository directory.
I ran several head-to-head and task-based comparisons between the internal engine and the neural model:
- Tactical puzzles (small composed tasks): Amateur players will quickly master these tasks - this is also reproduced by the internal engine, whereas the AI model was unable to do this reliably.
- Five personal matches(PGN): AI model 4 - Internal engine 1 (clear superiority of the AI in these encounters).
- The AI model learned several positional principles: piece development toward the center, space-gaining plans, and even material sacrifices to increase activity.
- The model showed a strong tendency to execute horizontal/linear back-rank mates when possible.
- The network is relatively small, and its combinational/calculation depth is limited - it struggles with long tactical variations. This weakness led to at least one game where the internal engine delivered an early mate via a simple combination.
Overall: the AI emphasizes positional play; the internal engine remains superior on short tactical puzzles and precise combination calculation.
flowchart TD
A["Input: Board(13×8×8) and Rating(1)"] --> B["Stem: Conv7×7 → BN → ReLU"]
B --> C[Residual Blocks: 96->128->128->192->192]
C --> X{Use Transformer?}
X -->|Yes| D["Flatten to 64x192 + Positional Embedding"]
D --> E["TinyTransformer: 2 layers, 8 heads"]
E --> F["Mean Pool -> Feature (192)"]
X -->|No| G[Global Avg Pool: 8×8 → 1×1]
G --> H["Flatten -> Feature (192)"]
F --> I[Concat Rating]
H --> I[Concat Rating]
I --> J["FC Common: (192+1->512) ReLU + Dropout"]
J --> K1[From-Head: 512->64]
J --> K2[To-Head: 512->64]
J --> K3[Value-Head: 512->64->1]
K1:::head
K2:::head
K3:::head
A:::head
classDef head fill:#0000f4,stroke:#333,stroke-width:1px,font-size:12px;
flowchart TD
A[Start Analysis]:::theme
A --> B[Depth = 1]:::theme
B --> E[Generate Moves]:::theme
E --> F[Alpha-Beta Search + TT]:::theme
F --> G{Quiescence at Leaf}:::theme
G -->|Yes| H[Capture-Only Search]:::theme
G -->|No| I[TT Probe/Eval]:::theme
H --> I
I --> J[Publish Analysis]:::theme
J --> L{Depth < Max}:::theme
L -->|Yes| E
L -->|No| M[Stop Analysis]:::theme
classDef theme font-size:12px