rogermt commited on
Commit
f31e44d
·
verified ·
1 Parent(s): e04ec87

Delete README.md

Browse files
Files changed (1) hide show
  1. README.md +0 -56
README.md DELETED
@@ -1,56 +0,0 @@
1
- # NeuroGolf 2026 Submission
2
-
3
- This repository contains solutions for the 2026 NeuroGolf Championship competition on Kaggle.
4
-
5
- ## Results
6
-
7
- - **Tasks solved:** 276/400 (69%)
8
- - **Total score:** 4106.7
9
- - **Average score per solved task:** 14.9
10
-
11
- ## Approach
12
-
13
- The solver uses multiple strategies in order of increasing complexity:
14
-
15
- 1. **Identity** - Output equals input
16
- 2. **Color map** - Per-pixel color remapping
17
- 3. **Transpose/Flip/Rotate** - Geometric transformations
18
- 4. **Tiling** - Conditional Kronecker product
19
- 5. **Upscale** - Nearest-neighbor upscaling
20
- 6. **Concatenation patterns** - Horizontal/vertical concatenation with transforms
21
- 7. **Position-based color LUT** - Per-position color lookup table
22
- 8. **Spatial gather** - Pixel rearrangement
23
- 9. **Linear/Affine** - Matrix multiplication
24
- 10. **One-hot convolution** - Encode colors as channels, apply convolution, decode via argmax
25
-
26
- The one-hot convolution approach proved most effective, solving 220 tasks. It encodes each pixel as a 10-channel one-hot vector, applies a learned convolution, and decodes via argmax.
27
-
28
- ## Solver Breakdown
29
-
30
- | Solver | Tasks |
31
- |--------|-------|
32
- | onehot_conv | 220 |
33
- | spatial_gather | 24 |
34
- | color_map | 4 |
35
- | concat_patterns | 4 |
36
- | pos_color_lut | 3 |
37
- | onehot_linear | 2 |
38
- | flip | 2 |
39
- | transpose | 2 |
40
- | upscale | 2 |
41
- | tiling | 1 |
42
- | tile_repeat | 1 |
43
-
44
- ## Files
45
-
46
- - `submission.zip` - ONNX files for 276 solved tasks
47
- - `solve_all.py` - Main solver script
48
- - `ARC-AGI/` - ARC-AGI training data
49
-
50
- ## Usage
51
-
52
- ```bash
53
- python3 solve_all.py
54
- ```
55
-
56
- This will create a `submission/` folder with ONNX files.