Files
foxhunt/docs
jgrusewski 97f69111c5 docs: DQN action collapse fix design — 5 fixes for 45-action diversity
Root causes: entropy regularizer dead code, noisy nets disable epsilon,
diversity penalty calibrated for 3 actions not 45.

Fixes: wire entropy into loss, fix normalization, epsilon floor for
noisy nets, sigma scheduling, UCB count-based bonus.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-03 00:59:20 +01:00
..