SirStone
62a6cc8ccf
fix(SAC_LSTM_Bot): training review fixes — hidden state ordering, redundant forwards, actor grad clip ( #47 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-21 00:16:47 +02:00
SirStone
415d4e3738
feat(SAC_LSTM_Bot): SAC training module ( #47 )
...
Implements sacUpdate with burn-in LSTM warm-up, twin-critic TD update,
actor reparameterization gradient, auto-alpha, and soft target update.
Manual backprop (linear + LSTM single-step, truncated BPTT). 11 new tests
all green; full regression suite (57+ tests) unaffected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-21 00:11:35 +02:00
SirStone
54b8139b11
feat(SAC_LSTM_Bot): weight persistence module ( #46 )
...
Save/load all SAC-LSTM tensors (actor, 2 critics, 2 target critics,
alpha, Adam states) into a single .zip of .npy files. Atomic write
via temp path + rename. Adam types (AdamVar, SACAdamStates) defined
here for training.nim to use.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:57:13 +02:00
SirStone
55ef22ff8b
Merge branch 'worktree-agent-a3ca3066' (ticket #45 replay buffer)
2026-08-20 23:41:23 +02:00
SirStone
5ea57bcae3
Merge branch 'worktree-agent-a393a8b9' (ticket #44 reward module)
2026-08-20 23:41:23 +02:00
SirStone
cb33551621
feat(SAC_LSTM_Bot): replay buffer module ( #45 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:40:17 +02:00
SirStone
23c65c9ac6
feat(SAC_LSTM_Bot): reward module ( #44 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:39:45 +02:00
SirStone
4ee0d8272c
feat(SAC_LSTM_Bot): action mapping module ( #43 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:38:45 +02:00
SirStone
add3e34926
feat(SAC_LSTM_Bot): LSTM network module ( #41 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:38:17 +02:00
SirStone
f130bf1254
feat(SAC_LSTM_Bot): state vector module ( #42 )
...
35-dim normalized tensor (GameState → buildState). No history window —
LSTM handles temporal context. Covers own-bot (7), enemy (7), derived (4),
walls (4), bullets (12), scan staleness (1). All tests pass.
2026-08-20 23:37:56 +02:00
SirStone
a0a3840980
feat(SAC_LSTM_Bot): skeleton bot with radar lock and colors ( #40 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:37:50 +02:00
SirStone
4258d364b9
feat(SAC_LSTM_Bot): project scaffold ( #38 )
...
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-08-20 23:33:43 +02:00