neurogolf-solver / medal-solvers /ANALYSIS_AGENT2.md
rogermt's picture
Add task062 solver (267/267 PASS) - reflection across axis, ONNX build pending
7cadfb0 verified
|
Raw
History Blame
1.97 kB

Agent 2 Analysis β€” Latest Status

Kaggle: 6065.98 (v42)

Task 062 β€” RULE CRACKED (267/267 PASS)

Rule: Input has a main-color shape and a color-2 axis (line or single cell). Reflect the entire combined shape (main + axis) across the boundary between the main shape's nearest edge and the axis. Output fills bg with color 3.

Key details:

  • Axis orientation: vertical if axis has 1 unique col AND main is to the side; horizontal otherwise
  • Mirror position: midpoint between the closest main-cell edge and the axis position
  • All cells (main + axis + reflected) become the main color in output
  • Background is color 3

ONNX viability: PROMISING

  • 10Γ—10 fixed grid β†’ slice-first
  • Base: 194 nodes, 929 params, score ~11.5
  • Target: ~60-80 nodes β†’ score ~14-15 β†’ gain +2.5 to +3.5 pts
  • Operations needed: ArgMax (find color), ReduceSum (find positions), arithmetic (mirror coords), comparisons (masks)

Python solver: 267/267 PASS (all train+test+arc-gen)

All Models Status

Task Gain Status Confidence
025 +1.81 βœ… On Kaggle Verified
028 +0.87 βœ… On Kaggle Verified
084 +2.56 βœ… On Kaggle Verified
153 +3.18 βœ… On Kaggle Verified
200 +3.56 βœ… On Kaggle Verified
255 +4.16 βœ… On Kaggle Verified
285 +2.83 βœ… On Kaggle Verified
319 +3.67 βœ… On Kaggle Verified
387 +0.47 βœ… On Kaggle Verified
062 +3 est Rule cracked, ONNX pending High
Total +26 est

Next: Build task062 ONNX

The ONNX implementation needs:

  1. Slice to 10Γ—10
  2. Find main color via ArgMax on channel sums (excluding bg and ch2)
  3. Find axis orientation using center-of-mass comparison
  4. Compute mirror position from main-cell boundary
  5. Reflect using 2*mirror - coord_grid
  6. Build output mask, apply colors, add bg, pad to 30Γ—30
  7. Apply strict shape inference fix