3bee4353b2e04729a1a2ed78828f7fb7e6c8a43e
5 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 SirStone
|
e40c8493a6 |
Offline harness: audited, calibrated against live, and one real bug fixed
AUDIT (docs/offline_harness_trust.md, new): - Re-ran acceptance_offline_vs_online myself TWICE: 12/12 deterministic guns exact both times (264 ticks/enemyId=1, 244 ticks/enemyId=2), death boundary included. The offline range reproduces the live bot's own per-gun virtual telemetry exactly. - Re-verified the (fireTick, powerBin) wave-pairing fix: exact-key lookup, collisions counted not silently mislabelled; test_wave_pairing 17/17 PASS. - The offline score is the live TELEMETRY (last-100 virtual hit rate) but NOT the live BATTLE score (damage/round wins). Two-level answer, documented. - bmPoint scores up to one tick-step (~17px) PAST its documented aim distance, while the tie-break probe scores exactly the aim point. Real, low-impact, deliberately NOT fixed (point metric is non-default, measured negative, and the committed point baselines would silently change). - bmPoint/bmPath, perfect-info captures, conditional-on-selection live rates, and hit-rate-as-objective-for-movement all catalogued as non-apples comparisons. FIX (unambiguous, fail-before/pass-after): - common_libs/tests/range_guns.nim: buildAllGunDrivers defaulted to enableTmSelector=true, so run_range / analyze_selector / test_power_selection / measure_power_policy spawned gun 13 (TMSelect) - a gun the shipped bot NEVER spawns. The shared VirtualTracker ring is order-sensitive, so those 4 spawns/tick permuted the learning guns' resolution order (the exact confound |
||
|
SirStone
|
4cd5618435 |
fix(test): repair the flaky offline==online acceptance; make the tie-break truly random
TASK 2 - THE FLAKY ACCEPTANCE TEST, root-caused. It was NOT a live/offline boundary race as suspected. The replay spawned gun 13 (TMSelect) while live has EnableTmSelector = false and never does. The shared VirtualTracker ring is ORDER-SENSITIVE, so gun 13's extra 4 bullets/tick shift the ring head and permute the per-tick RESOLUTION ORDER of every other gun. The learning guns append observations in resolution order, so their predictions shifted and produced small hit deltas that moved between runs. Evidence: the first KNN divergence was at rtick=174 with the SAME resolution set merely reordered (live ft133,138,139,142,148,150,151 vs offline ft150,151,133,138,139,142,148); after closing gun 13's ready gate offline the live and offline KNN traces became BYTE-IDENTICAL (diff empty, 904/904 lines). Fix: mirror the live rack in the replay. No tick exclusion, no tolerance loosening. Stability: 5/5 consecutive runs now report 12/12 exact, each with enemyDied=true - the death boundary is included, not excluded. The proof is now real rather than a lucky run. TASK 1 - the tie-break was not random. randomize() was only reached incidentally through initTsetlinGun(), so a rack without Tsetlin had a fixed rand() stream and ties always resolved the same way across process restarts. Added seedSelectorRng() after gun construction, honouring GUN_SELECTOR_SEED. Evidence: unseeded, 6 separate processes gave different pick sequences; with GUN_SELECTOR_SEED=42, 3 processes gave identical sequences. TASK 3 - PRUNING DOES NOT HELP; keep the full rack. 15 PAIRED runs per variant vs DrussGT, 8 rounds, identical seeds: baseline 3238 shots 6.18% (events 6.16%) 200 dmg/run Tsetlin disabled 3522 shots 5.76% (events 5.71%) 197 dmg/run Tsetlin+Displace 3478 shots 5.46% (events 5.37%) 183 dmg/run Paired permutation tests: -0.34pp p=0.57 and -0.70pp p=0.21. Per-run distributions completely overlap (baseline range [2.68, 10.00]; 15/15 and 14/15 runs inside it). A Crazy control showed no separation either. So removing the measured-worst real performers is neutral-to-slightly-negative, and with sd ~1.8pp a definitive claim either way would need far more runs. CORRECTION TO A CLAIM I MADE: the 'virtual metric is INVERTED' finding does NOT reproduce. Job-24 measured Spearman -0.374; this job measures +0.335 over the same 13 guns with a different but equally defensible aggregation. Two opposite signs means the correlation is NOT robustly negative - it is WEAK AND SIGN-UNSTABLE. The honest statement is that virtual hit rate is a poor ranker, not an inverted one. The docs assert the inversion and need correcting. Also adds per-process GUN_STATS_PATH/GUN_SHOTLOG_PATH so concurrent A/B runs do not clobber each other, and an env-gated GUN_RACK_DISABLE for rack A/Bs. All default behaviour is unchanged when the env vars are unset. |
||
|
SirStone
|
57b2ac3849 |
feat(guns): scale-aware power selection (+52% damage); TM classifier gun built, measured, DISABLED
TASK 2 - power selection, a clear win. bestPower used an ABSOLUTE MinHitRate = 0.40 bar. Measured per-bin virtual rates (rolling-100 fraction) show no bin ever clears 40%, so 11 of 14 guns were stuck at bin 0 (power 1.0) even where higher bins were comparable: Linear p1.0 44% p1.5 39% p2.0 30% p3.0 29% old bin 0 -> new bin 3 Accel p1.0 44% p1.5 40% p2.0 26% p3.0 29% old bin 1 -> new bin 3 Pattern p1.0 50% p1.5 40% p2.0 27% p3.0 12% old bin 1 -> new bin 2 Replaced with a scale-aware PowerBarFrac = 0.50 (a dimensionless FRACTION of the gun's own best bin rate). 13 of 14 selections now pick heavier bullets. Real effect vs DrussGT (8 rounds x 3 runs): hit rate unchanged (7.56% -> 7.47%) but damage dealt +52% (157 -> 239 per run) and rounds end faster. Same accuracy, half the shots, half again more damage. TASK 1 - the TM pattern-classifier gun does NOT earn its slot. It was built as a mixture of experts with a corrected-Granmo TM as a multi-class gate over HeadOn/Linear/Circular/WallBounce/Accel, labelled by which expert's prediction was closest to the actual enemy position (an exact, supervised, per-shot label - no delayed credit). Offline it loses to the best of its OWN experts on essentially every fixture, and against DrussGT it cost real performance: baseline (path+relative) 7.56% real hit rate, damage 157 + power fix 7.47%, damage 239 + power fix + TM gun 5.59%, damage 133 The gun was selected on 806 ticks and fired 24 real shots at 4.2%. So the tree ships with EnableTmSelector = false: code and wiring kept intact for re-enabling, but it is not in the active rack. Worth recording from the clause dump: the gate DOES latch onto meaningful structure. On energy-threshold-turner, HeadOn's clauses key on the energy bits (the rule's own driving variable) while Circular keys on distance/velocity. So the TM is learning something real and interpretable - it simply cannot beat 'always pick the best expert'. Root cause (INFERRED): the closest-expert label is noisy because several experts are near-tied, and under the path metric the winner varies by power bin while the gate sees one shared per-tick input, so a one-vs-rest gate over a saturated 870-bit clause space has no margin to exploit. (Zero-padding the 2-frame window was tried first and saturated every clause at 256-755 included literals; alternating the two real frames fixed that.) Also factors the corrected feedback into an exported tmLearnDir and exports the encoding/TM primitives; the Tsetlin tests still reproduce the documented mean=13.8 included literals, so the refactor is behaviour-preserving. Verified: 33/33 guard checks, tsetlin tests green, metric checks green, new power-selection guard green (13/14 selections change; relative bar still picks bin 1 and not bin 3 for a [30,25,12,5]% profile), 12/12 offline==online acceptance under the shipped default. |
||
|
SirStone
|
89370008da |
fix(tsetlin): make the TM actually learn - saturation 714 -> 13.8 literals/clause
The gun has never contributed anything: Tsetlin.vHits was byte-for-byte equal to Linear.vHits in every measured round of every run, because its learned correction was always exactly 0. Six diagnosed defects fixed, plus one that was required to make the first one work: 1. Type I now conditions on the clause output. It previously rewarded included true literals unconditionally, omitting Granmo's (c=0, lk=1) -> toward Exclude counter-force, so true literals ratcheted toward Include forever. This was the root cause of the saturation. 2. Type II was unreachable dead code: its guard required cOut==1 AND lits[lit]==0 AND st>0 (included), but cOut==1 guarantees every included literal is 1. Its direction was wrong too - it should increment EXCLUDED false literals when the clause fires. 3. Resource allocation restored: Granmo's (T - clip(v,-T,T))/(2T) target replaces |error|/(2*RESID_MAX); TM_T was only an output normaliser. 4. Label baseline fixed - the factor-2 shrink. predX = linearX + cx, so the label was delta - cx while the learner's output IS cx, giving error = delta - 2cx and a fixed point of cx = delta/2: HALF the needed correction even with perfect feedback. TmTrace now stores linearX/linearY and training uses delta. 5. Hits no longer zero their label (a hit means |miss| < 18px, not 0). 6. The enemy-energy feature was duplicated - tmEncodeFrame passed state.selfEnergy with a stale comment claiming enemyEnergy was absent, while WorldState.enemyEnergy exists. Enemy-energy rules were literally unrepresentable. 7. REQUIRED EXTRA: tmEvalClause now implements Granmo Eq. 6 - an all-Exclude clause outputs 1 during learning and 0 during classification. Without it, fix #1 deadlocks every clause at empty. MEASURED EFFECT (energy-threshold-turner fixture, seed 1): mean included literals per active clause 714.0 -> 13.8 active clauses 100/100 -> 53/100 nonzero corrections 8/764 -> 708/764 Tsetlin virtual hits (Linear = 27/400) 27/400 -> 69/400 Divergence achieved: offline on 7/8 fixtures, and in a live gauntlet (RandomMover: Tsetlin 199/1200 vs Linear 288/1200, vDropped=vStarved=0). Tsetlin now LEARNS but is not yet competitive with Linear - the regression head is untuned, flagged as follow-up rather than claimed as a win. Also ignores compiled test harnesses that have no file extension, which the existing '**/tests/test_*' rule misses. |
||
|
SirStone
|
974528d5cf |
feat(gun_harness): offline gun range, proven equivalent to live play
Gun evaluation previously required a full end-to-end battle (Java server + battle runner + websocket IPC to 2 bot processes, 50 rounds, ~3.4 min) and yielded only ~300-900 REAL shots across 13 guns -- far too few to rank guns, which is why tuning needed many repetitions. VirtualTracker is already a pure function of (WorldState stream, gun list); the only reason it needed Java was where WorldState came from. So the range replays a seq[WorldState] through the SAME tracker: offline and online scores are the same metric by construction, not an approximation. ACCEPTANCE TEST (the point of the whole thing): record one live round, replay it offline, compare per-gun virtual hit rates. 12/12 deterministic guns match EXACTLY, reproduced twice. Tsetlin is compared separately because tmLearnOne calls rand(). Getting to 12/12 exposed two real ordering quirks in the live loop: run() calls go() before the aim/fire block, so tickBullets resolves against the NEXT tick's scan while the prediction used the previous one; and if the target dies during that go() the final tick's spawn+resolution is skipped entirely. The recorder emits an end marker for the second case. The 5th (selected-gun) predict call was verified to be a no-op. Measured cost: 8 fixtures (1770 ticks, ~92k virtual bullets, 13 guns) replay in 2.9 s, ~32k virtual bullets/s -- roughly 70x faster and 100x more samples than a live gauntlet. Also adds a per-tick WorldState recorder behind const RecordWorldState (default off, mirrors the ShotLog idiom) which records the state the bot ACTUALLY builds, staleness included, rather than true positions -- recording the latter would hand the guns perfect information and produce flattering scores. 9 new guard checks (33 total, all passing), including fixture round-trip, replay determinism, stationary->HeadOn 100%, constant-velocity->Linear>HeadOn, and the energy-threshold turner crossing at t=41. |