sac_train.sh orchestrates chunked self-play via tools/training_runner/
RunTraining.java: weighted opponent sampling per chunk, deterministic
eval (SACLSTM_EVAL_MODE=1) every N chunks with win-rate tracking, best
checkpoint (weights/sac_best.zip) by eval score, crash-restart loop on
the runner's liveness detection.
Supporting changes:
- integration.nim: opponentKey() keys the NewBattle buffer-clear rule on
getBotName(id) with numeric-id fallback (#49 Q14 follow-up);
bumpRoundCounter() emits the per-round liveness signal.
- SAC_LSTM_Bot.nim: onRoundEnded -> bumpRoundCounter().
- RunTraining.java: BOT_NAME env parameterizes result matching
(default PPO_Bot, unchanged behavior for PPO).
- Launch packaging: root SAC_LSTM_Bot.json + .sh for the booter;
src json name aligned to 'SAC_LSTM_Bot' so self-reported identity
matches the booted identity (mismatch = runner connect timeout).
Stochastic eval at std≈0.37 was 0/10 vs Corners (deterministic: 10/10).
Warm-start policy is correct but brittle — any noise breaks it.
- log_std initialized to -2.0 (std≈0.135) for moderate exploration
- entropy_coeff=0.0 (no push toward exploration during fine-tuning)
- logStd ceiling=-1.0 (cap at std≈0.37)
- Accumulate transitions across 10 rounds (~3000) before PPO update
(was per-round ~300 — gradient estimates were far too noisy)
- training.nim: MAX_TRANSITIONS 4096→8192, done flag on transitions,
GAE handles episode boundaries correctly
- PPO_Bot.nim: buffer persists across rounds, update every N rounds
- training.env: lr 5e-5→1e-4, entropy 0.001, UPDATE_INTERVAL=10
- Bullet state (indices 44-55): enemy-relative → bot-relative frame
(bot needs threat vectors to itself for dodging, not to enemy)
- New index 56: scan staleness = min(ticksSinceLastScan / 30, 1.0)
(gives policy a confidence signal for enemy data freshness)
- warm_start.py updated: 44→57 dim expansion, TARGET_DIM variable
- Tests updated for new state layout
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
computeRoundReward used cumulative totalScore/50 — unbounded in long battles
(vLoss 353 at round 3160 → 25745 by 3871 in the 5841-round attempt). Cap the
score term at 400 before /50: bonus ∈ [0,8], so the critic's value scale stays
stable regardless of battle length and across battle boundaries.
Reverts the 60-round battle chunking (186e005/da2f825): one battle per
campaign for the whole remaining budget; keeps the crash-restart loop, the
mid-battle freeze guard and the end-of-battle counter completeness check.
Cert (5211, single 60-round battle): 59/60 wins (sole loss = cold-start round
1, score 61), vLoss avg 36.1 / max 148.5, gNorm max 596, zero NaN, zero
restarts, counter check passed. Weights persist to round 5211.
RunTraining exits 0 after each chunk; && break ended the whole run after
the first battle (counter 3219, not 9000). Loop now falls through the
success path and re-checks the persisted counter each iteration.
Round-end reward = cumulative totalScore/50 grows unboundedly with battle
length; long battles (5841 rounds) blew the critic's value scale: vLoss
10-30 during the 60-round cert, 353 at battle-1 round 1, 25745 by round 3871,
policy drift to 0/6 wins. 60-round battles reproduce the certified regime:
bounded value targets, fresh bot process per battle (clears thread state).
The event queue's heap seq was the last GC'd block surviving across
rounds: each round runs on a freshly spawned bot thread, so the N+1
thread realloc'd a block grown by dead thread N's allocator mid-round
(at the next capacity doubling, ~turn 104) -> rawDealloc SIGSEGV in
addEvent (7 gdb-confirmed coredumps). Replace with a static
array[MAX_QUEUE_SIZE, BotEvent] + eventsLen: no heap block crosses
threads, realloc can never happen.
Also fix the harness aborting the final round mid-train: PPO_Bot's
onRoundEnded trains synchronously after the runner's RoundEndedEvent,
so the counter read right after awaitResults() is the stale pre-train
value and System.exit killed the bot inside ppoUpdate. Poll up to 60s
for the counter to catch up before declaring the battle incomplete.
Verified: 72 consecutive rounds vs Fire, 100% wins, all rounds trained
(counter advanced 1:1), zero coredumps since the fix.
SirStone