From 6fa3505e1220ff4ea4b49d4d916b41164d150314 Mon Sep 17 00:00:00 2001 From: jgrusewski Date: Wed, 8 Apr 2026 01:12:24 +0200 Subject: [PATCH] =?UTF-8?q?fix(v8):=20increase=20micro=5Freward=5Fscale=20?= =?UTF-8?q?0.001=E2=86=920.01=20(10x=20stronger=20bootstrap=20signal)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit At 0.001, the dense micro-reward was 1/1000th the magnitude of the sparse exit reward (±10). Too weak to meaningfully bootstrap Q-values. At 0.01, it's 1/100th — strong enough to provide directional feedback while still subordinate to the exit signal. Co-Authored-By: Claude Opus 4.6 (1M context) --- crates/ml/src/trainers/dqn/config.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/crates/ml/src/trainers/dqn/config.rs b/crates/ml/src/trainers/dqn/config.rs index 633325fa3..4791e8b0c 100644 --- a/crates/ml/src/trainers/dqn/config.rs +++ b/crates/ml/src/trainers/dqn/config.rs @@ -1656,7 +1656,7 @@ impl DQNHyperparameters { exposure_aux_weight: 0.5, exposure_aux_warmup_epochs: 5, // v8: Dense micro-reward, TD(λ), PopArt, curriculum, hindsight - micro_reward_scale: 0.001, + micro_reward_scale: 0.01, td_lambda: 0.9, max_trace_length: 7, popart_enabled: false, // disabled by default until wired