mrvictoru/energydecision-dt-grpo

AEMO Decision Transformer — GRPO Post-Trained (Legacy 8x384)

⚠️ OUTDATED — Superseded by the Stage C standalone DT.

This GRPO fine-tuned checkpoint is the legacy 8×384 DT. It was already superseded by the modern v2 pretrained model (no GRPO needed), which in turn was superseded by the Stage C standalone DT distilled from an honest SDP-planning teacher:

Surface	Stage C DT	This model (GRPO)	PPO
Standard Oct	$11,573	$2,678	$2,353
Dispatch-matched	$35,320	$8,242*	$22,530
Expanded broad-2024	$34,761	$2,459	$19,504

* Overfit — collapses to $1,533 on the standard surface.

This checkpoint is retained for reproducing the legacy GRPO study only.

---

Overfitting caveat

The headline dispatch-matched number ($8,242) is a specialist overfit to the SA1 Q4 2024 benchmark. On the broader standard surface this is the worst cross-region generalization of any DT model — worse than even the legacy pretrained baseline ($2,459).

Recommendation: Use the Stage C standalone DT (models/aemo/dt/aemo_dt_sdp_jtsoc_fullcorpus.pt) for any new work.

GRPO Configuration (Phase 1)

Parameter	Value
Iterations	5
Learning rate	1e-5
KL coefficient	0.02
Group size	8
Training regions	NSW1, SA1, QLD1, VIC1, TAS1
Training date range	2024-01-01 → 2024-09-30