Fix R-GSP-N: LSTM hidden state + vectorized learning + stored state (R2D2)#20

Merged

jdbloom merged 4 commits into

mainfrom

feat/rddpg-lstm-fix

Apr 9, 2026

Commits on Apr 8, 2026

fix(lstm): propagate LSTM hidden state through EnvironmentEncoder and RDDPG networks

Joshua Bloom
and
claude
committed
feat(buffer): store LSTM hidden state at sequence boundaries (R2D2-style)

Joshua Bloom
and
claude
committed
feat(rddpg): vectorize learn_RDDPG with burn-in for 10x speedup

Joshua Bloom
and
claude
committed
feat(device): remove MPS CPU fallback for recurrent — vectorized RDDPG works on MPS (80ms vs 150ms CPU)
Joshua Bloom
committed