SirStone/SirRoboGarage
1
0
Fork 0
Code Issues 28 Pull Requests Actions Projects Releases Wiki Activity

fix(SNNBot): increase learning rates — ETA 0.05→0.1, ETA_IH 0.005→0.02

Browse Source 
Weights were barely moving: |wih| +0.14%, |wOut| +7.7% over full session.
Bump output ETA 2x and input→hidden ETA 4x to accelerate convergence.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Davide Cappellini
2026-09-14 22:21:29 +02:00
parent 80dbf81479
commit a80fd8d656
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
SNNBot_garage/src/SNNBot.nim
+2 -2
View File
@@ -30,8 +30,8 @@ const
THRESH = 0.08 # ponytail: THRESH=0.08 tuned for 80-input layer; raise if firing saturates
MAX_GUN_TURN = 20.0 # max gun turn per tick (degrees)
AIM_TOL = 2.0 # arrive tolerance (degrees)
ETA = 0.05 # SuperSpike learning rate for hidden→output weights (r_0 from paper)
ETA_IH = 0.005 # ponytail: ETA_IH=0.005 scaled down for 80 inputs; raise if input→hidden converges too slowly
ETA = 0.1 # SuperSpike learning rate for hidden→output weights (r_0 from paper, bumped 2x)
ETA_IH = 0.02 # ponytail: ETA_IH=0.02 bumped 4x for input→hidden (was 0.005, too conservative)
# ponytail: separate input→hidden rate; add RMaxProp optimizer if convergence still unstable
N_INFER = 10 # inference window ticks per DECIDE
# ponytail: N_INFER=10, increase if output still noisy; decrease if too slow per tick