jgrusewski 909650f2a1 plan: three-layer stable distributional loss architecture ...

Layer 1: Per-sample C51 loss clamp (MAX_CE=50) breaks PER feedback loop
Layer 2: Label smoothing (ε=0.01) on Bellman target prevents log(0)
Layer 3: IQN primary with fixed τ (QR-DQN style, CUDA Graph compatible)

5 tasks, each independently testable and committable.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-23 20:42:08 +01:00

..

infra

feat(infra): GPU training launcher with local/cloud routing

2026-02-23 11:19:58 +01:00

plans

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

superpowers

plan: three-layer stable distributional loss architecture

2026-03-23 20:42:08 +01:00