6fd5fe632893c125c629d5f39f100b91c749330d
368 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 SirStone
|
2747ebd323 |
BitBrain: TR_BITBRAIN_GAINS env knob (candidate set + fixed-gain degenerate)
Task A of campaign phase 2: the lead-gain candidate set is now pure env, so the live arms need no recompile. - common_libs/guns/bitbrain_gun.nim: BB_GAINS_ENV (TR_BITBRAIN_GAINS); the candidate list is parsed once at gun construction into a dynamic seq, so the hit counts/hit rates are sized to it. Unset/unparsable -> the shipped BB_CAND set [0,0.25,0.5,0.75,1.0] (byte-identical behaviour). Exactly ONE candidate degenerates to a FIXED gain applied from the first shot (learning bypassed), still gated to the long bands. parseGains clamps to [0,8], de-dupes and sorts so the argmax tie rule is unchanged. The [bb] line now prints the APPLIED gain AND the resulting angular shift, so a run's correction is auditable from stdout. - ModularBot_garage/src/env_report.nim: emit TR_BITBRAIN_GAINS (resolved candidate set) and add BB_GAINS_ENV to the known-name list. - tools/ab/arms_leadgain.txt: the 6-arm phase-2 sweep definition. |
||
|
SirStone
|
c305ef4212 |
BitBrain campaign phase 1: the missing gain sweep and BitBrain as a lead-gain corrector
Task A — the gain region Phase 0 never covered (gain < 1). Extend the prediction-quality ruler with gain 0.25/0.50/0.75 arms and a fixed causal per-band arm. Full 70-run result: the hitProxy-argmax curve is [1.00, 1.00, 1.00, 0.00, 0.00] — Pattern below 300 px, HeadOn above — worth +0.13 pp at 300-450 and +2.16 pp at 450+ (0.0767 -> 0.0984). Lead correlation is identical for every g>0 (Pearson is scale-invariant), so a shrinking gain adds no lead information; and the least-squares optimum [1,1,1,1,0.25,0.25] diverges from the hitProxy optimum because Pattern's lead errors are bimodal. Task B — rebuild guns/bitbrain_gun.nim as a lead-gain corrector: aim = LOS + gain*(patternAim - LOS), gain learned online per range band by ranking candidate gains on the hit-probability proxy (the observed lead label via tmhObservedAt), gated to range >= 300 px. ADE+SBC output removed. Offline (70 runs): Pattern below 300 px, +0.37 pp at 300-450, +1.90 pp at 450+ (hitProxy 0.0957 vs 0.0767), matching the fixed rule to within 0.27 pp at 450+ and exceeding it at 300-450. ~0.0007 ms/tick marginal (old gun ~0.114 ms/tick). Default off; rack membership, env report and guard tests (bitbrain 32, registration 13, rack 48, tm_pattern 20, env_report) unchanged and green. Ledger: docs/bitbrain_campaign.md Phase 1, including the three on-file negatives and the causal-shippability note. No live claim. |
||
|
SirStone
|
140fe2519a |
HeadOn (no-lead) vs Pattern LIVE at long range: clean negative, offline ruler killed
2 arms x 15 runs x 7 rounds, one frozen binary from HEAD
|
||
|
SirStone
|
a82c864c60 |
bitbrain campaign phase 0: offline prediction-quality ruler and the bar
New harness (common_libs/gun_harness/prediction_quality.nim + common_libs/tests/run_prediction_quality.nim): per-gun single-tick aim error in degrees against the true continuous interception point on the recorded live-vs-real-DrussGT corpus (/tmp/tfil_ab2/out, 70 runs, 899607 ticks), per range band, with the hit-probability proxy mean(|err|<=atan(18/range)). Validated: recorded hits separate from misses 13.34x px (reference 11.59x), perfect-oracle max |err| = 0, correct ordering on synthetic ground truth, two full runs byte-identical. Fixed a wrap180 bug (Nim float mod keeps the dividend sign) that inflated the negative error tail. Bar (mean|err| deg [hitProxy] at 450+): Pattern 16.19 [0.077], naive-linear 22.86 [0.054], TMHorizon 16.20 [0.076], BitBrain 16.20 [0.077], static HeadOn 12.33 [0.098], oracle 0 [1.0]. Lead-gain sweep on Pattern is a dead end (1.0 wins every band). Naive-linear applies ~1.8x Pattern's lead but carries no more lead information (corr 0.178 vs 0.165) and is strictly worse. Ledger: docs/bitbrain_campaign.md. All verdicts remain live-only. |
||
|
SirStone
|
32a5e72fac |
Gun mixing (TMHorizon+BitBrain) vs DrussGT: clean negative, no dodge disruption
4 arms x 7 runs vs real DrussGT. mix alternates the two guns 476 times/7 runs (liveness OK) but our bullets are no more varied (power sd / aim-offset sd flat) and DrussGT's dodge quality is unchanged (miss/tick mix-pat +0.03, p=0.66; MDE 3.8%). mix wins 24/49 = the 49% baseline; the user's 6/10 has P=0.353 at 49%. New tools/ab/ab_dodge_analyze.py splits the validated per-shot dodge instrument by arm and adds gun-switch/power/bearing liveness; fixtures committed. |
||
|
SirStone
|
d93ce444c0 |
BitBrain verdict: clean negative at 30 runs/arm; analyzer gets MC + Mann-Whitney + MDE
- docs/bitbrain_gun_verdict.md: control vs bb_decay (decay SBC memory) vs a provably-zero placebo, 30 runs/arm vs real DrussGT. Nothing separates (bb_decay +3.3 dmg/run, p=0.71; round wins 97/210 vs 97/210, p=1.00); the 7-run shape does not replicate. TR_BITBRAIN_RANGE=0 is clamped to 1.0 deg (bitbrain_gun.nim:207) so it is NOT a zero-shift placebo; TR_BITBRAIN_MIN_OBS unreachable is used instead. - tools/ab/ab_analyze.py: keep exact enumeration for C(n,na)<=20e6 (7v7), add a seeded Monte-Carlo permutation test (1e6 draws, 0x5eed5eed) with its standard error, a tie-corrected Mann-Whitney U cross-check, a minimum detectable effect line, all-pairs comparisons, and a [bb] shift check. - tools/ab/README.md: document the new analyzer output. |
||
|
SirStone
|
f91e121965 |
lead capture by range: we under-lead (0.40->0.135) — a RANGE effect, not a power one
Measure, for every shot ModularBot fires at the real DrussGT, the lead we actually applied vs the lead the enemy's motion required, from the recorded live battles (/tmp/tfil_ab2, 70 battles / 490 rounds / 54926 shots, plus a 35-battle powtest replication of a different binary). - requiredLead from an AIM-INDEPENDENT interception solve (bullet speed vs enemy truth), appliedLead from the server-recorded bullet bearing. - capture = applied/required, guarded at 2px lateral lead (1.6% excluded); headline metric is the robust proportional slope. - validation: hits 11.6px mean miss / 80.8% inside 18px, misses 134px, 11.6x separation; 496/496 death + 70/70 owner attributions correct. Direct answer: capture falls with RANGE (capSlp 0.401 -> 0.135, and |err|/tolerance 1.27 -> 7.54) but is FLAT across fired POWER within a band (450+, enemy alive: 0.154 / 0.127 / 0.127). The sub-0.5 long-range shots (1.18% hit) are finishKill endgame shots at a near-dead DrussGT, not a lead-capture failure. A naive linear predictor captures 0.29-0.60; we reach 46-67% of that, so the under-lead is real but capture=1.0 is unattainable against a dodger (oracle required lead). |
||
|
SirStone
|
795a0e59fe |
BitBrain gun (id 16): Pattern-relative ADE+SBC aim corrector, default off
Wire the verified common_libs/bitbrain ADE+SBC library into ModularBot as a
fine-grained angular corrector on top of Pattern's prediction, the shape the
offline gate test measured (argmax readout over N correction classes).
- common_libs/guns/bitbrain_gun.nim: new gun. Input = the existing TMHorizon
53 bits (tmhBaseBits + tmhLits); output = argmax class centre over
+-TR_BITBRAIN_RANGE, applied by rotating the Pattern point around the shooter
exactly as tmhApplyShift does. Label = the +h-tick fact from TmHorizonGun's
own observation ring (never across a round). Prequential (defer + resolve).
AD layer synthesised online for our binary inputs (center=0): heuristic
cold-start thresholds + running-histogram ~1% percentile init + the library's
adaptThresholds. Memory modes perRound (default, measured best) / retained /
decay (periodic partial SBC wipe). Lazy network build + local RNG, so the
default path builds nothing and consumes no global randomness.
- tm_horizon.nim: export tmhUpdateHistory and add tmhObservedAt (label seam).
- selector.nim: register BITBRAIN at rack id 16, default rmOff, in the SAME
commit as the id and the wiring (the
|
||
|
SirStone
|
1fec87537d | gitignore: keep the BitBrain primary-source zip local (explicit path, not a pattern) | ||
|
SirStone
|
17c50159ba | tools/ab: reusable A/B runner + analyzer (frozen-HEAD build, exact permutation test, liveness check) | ||
|
SirStone
|
e670788eee |
env_reference: the ring mover was NEVER measured offline - correct a false label
The offline-harness audit (`e40c849`, `docs/offline_harness_trust.md`) found that the claim "best offline hit rate of anything measured" for the ring mover was false. The 20.28% figure is a LIVE number: `docs/feature_ab_results.md` and commit `bfdcdf8` record 35 real-DrussGT bridge battles with a server-side event sidecar as the ground truth, and the 6/49 round wins is likewise live. There is no offline measurement of the ring mover anywhere - the offline harness scores GUNS, not movements, and has no movement driver at all. So this was NOT an offline-vs-live calibration failure, which is how it has been described repeatedly (including by the orchestrator). It was a METRIC MISMATCH: a movement arm judged on hit rate instead of damage/run and round wins - and hit rate is precisely the metric that concealed its collapse. The lesson previously attached to this result was therefore the wrong one. The paragraph now says what actually happened and points at the audit. Docs-only; no code touched. |
||
|
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
|
f41cd08718 | BitBrain gate test: fine-grained aim correction vs naive + Pattern (offline) | ||
|
SirStone
|
1adefaba26 |
DrussGT dodge vs fired power: no movement response once range is controlled
Answers the user's hypothesis that DrussGT dodges low-power shots better. Measured on 70 live battles / 490 rounds / 54939 real shots vs real DrussGT (/tmp/tfil_ab2) and replicated on 35 more battles / 24280 shots (/tmp/powtest). Power is not randomly assigned - our policy caps it by RANGE (TR_POWER_FAR_DIST=200 -> 1.0) and by OUR OWN ENERGY (the slope), so inside a range band power is almost a deterministic function of our energy and a naive low-vs-high comparison is secretly a losing-vs-healthy comparison. Everything is stratified by range band and backed by a within-band shuffled-label null (arrival re-derived, so the null keeps the kinematic channel), a round-cluster bootstrap, and a within-shot CONTROL window 40 ticks later when the bullet is long gone. RESULT: no behavioural response. In band 450+ the raw miss distance at arrival is +8.25 px [+5.39,+11.25] for HIGH power - but per flight tick it is 4.52 vs 4.51 px/tick (delta -0.01 [-0.12,+0.10]), i.e. entirely the 2.13-tick longer flight window of the slower bullet. Fixed-12-tick lateral displacement is flat (55.63 vs 55.47, -0.15 [-1.15,+0.84]) and turn rate / speed are flat. The whole difference is already present 5 ticks after the trigger pull (+4.2 px) and is just as large in the bullet-free control window (+5.6 px), so it is a property of the low-energy situation, not of the shot. Hit rate is flat (0.10 vs 0.09). Corpus/attribution notes: e*=DrussGT (subject), s*=ModularBot, per TrBattleCapture.java; the Tank-Royale owner id is NOT stable across runs and is recovered per battle from fire geometry + the energy decrement, cross-checked on 496/496 death events. Geometry validated on the server's own hits (mean miss 11.6 px, 80.6% inside the 18 px radius). |
||
|
SirStone
|
77e6dace01 |
BitBrain: generic clean-room ADE + SBC library with MNIST acceptance
Implement the BitBrain (Address Decoder Element + Sparse Binary Coincidence) classifier as a generic, deterministic Nim library under common_libs/bitbrain/, written from the published algorithm (Front. Neuroinform. 17:1125844), not from the GPL-3.0 reference C. - ade.nim: signed thresholded random projection (scale 64 / centre 127 defaults reproduce the reference), multi-width ADs, optional deterministic homeostatic threshold adaptation. Hebbian longevity and Metropolis-Hastings sampling are described but not implemented. - sbc.nim: packed class-bit coincidence memory; idempotent learn, counting inference. - bitbrain.nim: container over several ADs and SBCs, online learn/infer, argmax readout, memory accounting. - tests: 32 unit checks (idempotence, planted rule + monotone online curve, shuffled-label chance control, unseen input, homeostasis, memory). - tests/test_bitbrain_mnist.nim: loads the reference pretrained ADs/thresholds and MNIST from /tmp, reproduces the reference exactly - 97.210% corrected and 96.540% bug-compatible - confirming the port. No gun/wiring integration yet; inputs and outputs to be agreed separately. |
||
|
SirStone
|
cc11ede824 |
TFIL: measure the commitment A/B on 490 live DrussGT rounds - no effect
Runs the five-arm commitment A/B that |
||
|
SirStone
|
c8b2a8c6c5 |
docs: reverse-engineer the BitBrain algorithm from its C source; assess as a gun
Reads docs/BitBrain_C_code.zip end to end (full_mnist_2048.c 568 lines + bitarray.h), reconciles every weight/data file's byte size against its loader, and answers two questions: 1. What the algorithm is: signed thresholded address decoders (random projections) -> write-once sparse binary coincidence bit tensors -> counting/argmax readout. Thresholds and ADs come from an unsupervised program that is NOT in the zip; only the SBC tensors are learned here. The supervised rule is genuinely online, single-pass, order-free, and has no learning rate - so it satisfies ModularBot's reset-on-enemy-change constraint. Found and isolated a real bug: read_from_sbc's uint8_t bit_test truncates the 32-bit bit test, so only ADEs with i%%32 < 8 are ever counted. Shipped reader: 96.540%% (reproduced exactly); with the bug fixed: 97.210%%. 2. Whether it can become a gun: not on this evidence. Measured full inference at ~0.56 ms/sample (cost is not the blocker), but the AD layer is unobtainable, and this repo has already shown the binding constraint is the target signal, not the learner - the TM head scored 2.08pp below its own majority class and the AD/SBC primitives are already in the 16-family shootout as WiSARD/Bloom/SDM. Recommends a readout swap on the existing WiSARD feature extractor and a Gate-2b-style >=80%% side-accuracy gate before any port. Also: recommends committing the 11 MB zip under its explicit name. |
||
|
SirStone
|
19bf4610e4 |
TFIL: env-gated commitment arms (tile-replan cancel is a bug)
The tile-change replan cancels the 15-tick movement commitment whenever OUR
tile changes. With GridSize=36 and speed up to 8 px/tick that is every ~5
ticks, so the commitment is cancelled by the motion it commands (measured:
96.9% of picks were tile-change replans, 33.8% of picks reversed direction).
Adds four env knobs, every default reproducing the shipped mover
byte-for-byte:
TR_TFIL_TILE_REPLAN self (default) | off | enemy
TR_TFIL_COMMIT_TICKS 15 (default)
TR_TFIL_NO_REV 0 (default)
TR_TFIL_COMMIT_LOG off (default, JSONL per-tick diagnostics)
- `off` honours the commitment; the danger replan stays the safety valve.
- `enemy` keys the cancel to the TARGET's tile displacement (the intent the
original comment claimed).
- `TR_TFIL_NO_REV` down-weights (never filters) tiles >90 deg from the travel
direction; the pool can never be emptied.
Default-path parity is guarded by test_tfil_commit_env.nim, which replays
tools/fixtures/tr_drussgt_vs_modularbot.jsonl and diffs every move command
against a golden generated from the pre-change build (git archive
|
||
|
SirStone
|
f842ac0f76 |
Env report Part 2: warn on misnamed env vars (unknown + compile-time defines)
The boot report printed the raw env and the resolved values, but it never
told the user when a name was WRONG - which is the failure mode that cost
real time: `TMH_NSTATES` was exported as an env var although it is a
compile-time `{.intdefine.}` (`guns/tm_horizon.nim:102`), so the export was
a silent no-op. Add the two warning paths, both boot-only and stdout:
* unknown TR_*/GUN_* names are named explicitly. The known set is built
from the modules' exported env-name constants; the remaining inline
reads are listed once, and common_libs/tests/test_env_report.nim scans
the tree and fails if a name read anywhere is missing.
* any `{.intdefine.}`/`{.strdefine.}`/`{.booldefine.}` symbol present as
an env var is flagged, with the real runtime equivalent when one exists
(`TMH_NSTATES` -> `TR_TMHORIZON_NSTATES`) and an explicit "does not
exist - this knob is compile-time only" when it does not. The map is
derived by grepping the tree; the same test re-greps and fails on drift.
Warnings are emitted only when the env is dirty, so a correct run keeps the
documented A/B/build shape. TR_ENV_REPORT=0 still suppresses everything.
test_env_report.nim: 25 checks (known set, pure helpers, tree literal scan,
tree define scan). All 15 existing guard tests keep their exact counts.
|
||
|
SirStone
|
036979e78e |
env_reference: verify every knob against the code, fix the trap that cost real time
DOCS ONLY. The user set TMH_NSTATES, which is a compile-time {-d:intdefine.}
(-d:TMH_NSTATES=2, tm_horizon.nim:102), not an env var; the runtime var is
TR_TMHORIZON_NSTATES (tm_horizon.nim:139). This rewrites the reference so that
class of confusion cannot recur.
What was WRONG and is now fixed:
- "unparseable warns and falls back" was false for the numeric/bool knobs: only
the enumerated string knobs warn; envInt/envFloat/envBool fall back silently.
- TR_POWER_LOG/TR_RAM_LOG/TR_MOVEMENT_LOG/TR_RECORD_WORLDSTATE/
TR_RADAR_FORCE_SPIN/TR_RADAR_SCANLOG/TR_TRACKER_PROBE are read with existsEnv,
so TR_POWER_LOG=0 turns the log ON. Documented per knob.
- TR_POWER_ENERGY_MIN is a CAP at low energy, not a minimum-power floor.
- TR_TMHORIZON_* knobs are inert unless TR_RACK_TMHORIZON=both; the doc implied
they were live.
- GUN_SELECTOR_WINDOW is clamped 1..100; TR_MOVEMENT silently falls back to tfil
for any value other than tfil_ring.
- The ring file's own header comment (corridor 5 / wall 10) is stale; the code
defaults are 10.0/15.0 (commit
|
||
|
SirStone
|
9bf3005850 |
Boot-time env report + fix the env reference
The bot is spawned by the server/GUI, so it inherits the SERVER's environment. The user could not tell whether their exports reached the bot, so print a one-shot greppable report at boot: grep '^\[env\]' /tmp/modularbot_stdout.log Section A prints every TR_*/GUN_* this process actually received, the count vs the total env size, a loud warning when nothing matched, and the process identity (pid/ppid, cwd, self command line, and the PARENT command line) so the spawn trap is obvious. Section B prints the resolved effective value of every documented knob with its source (env|default), including clamps and the rack's empty-set fallback. Build identity (NimVersion, compile date/time, binary path/size/mtime) pins the exact artifact. Suppress with TR_ENV_REPORT=0. docs/env_reference.md: add the missing GUN_SHOTLOG_PATH, GUN_SELECTOR_MINOBS/FLOOR/POOL/RANK/SHRINK/SEED, TR_ENV_REPORT and -d:TM_NCLAUSES; record the measured TR_TMHORIZON_WINDOW verdict; and add a prominent 'Did my env vars actually reach the bot?' section with the boot report, the /proc/PID/environ no-code check, the correct GUI launch recipe, and how to prove the trap deliberately. |
||
|
SirStone
|
7f6ccfb015 |
Ring mover heat retune (author: the user) - dodge bullets, ignore walls/centre,
re-plan 3x more often Uncommitted working-tree change in `the_floor_is_lava_ring.nim`, confirmed by the user as theirs. Committing it so it stops appearing in every job's `git status`. WHAT IT DOES - a coherent "be far more afraid of bullets" strategy: - BULLETS 2x hotter: BulletCore 10 -> 20, BulletAura 5 -> 10. - Walls weaker and thinner: WallRadiance 10 -> 5; the `WallHotness` default 10 -> 15 (so the heat falls off over ~3 tiles instead of 1, with a higher peak). - CENTRE PILLAR DISABLED: PillarHotness 30 -> 0, PillarRadiance 10 -> 0, so the bot may now use the middle of the arena instead of treating it as a no-go zone. - MUCH more reactive: CommitTicks 15 -> 5 and MinCommitTicks 5 -> 0, i.e. the dodge target is re-chosen 3x more often and a replan is allowed immediately. - `CorridorHeat` default 5 -> 10. This applies ONLY to the ring mover, which is NOT the shipped movement (the default is `tfil`), so the shipped bot is unaffected. TWO FACTS RECORDED, not objections - it is the user's call: 1. `CorridorHeat = 10` now EQUALS `PathDangerThreshold = 10`. The earlier measured taming set it to 5 precisely so a single corridor could no longer poison a path on its own (a corridor at 10 puts every tile along it at/over the threshold). That is part of what lifted the "band-weightable" fraction from 10.8% to 26.5%. At 10 the range weighting gets less to work with. 2. UNMEASURED: this retune has not been A/B'd. The ring mover it modifies was itself measured as a glass cannon (best offline hit rate, but round wins 16/49 -> 6/49, p=0.012). The two changes push in the same direction - hotter bullets and more frequent replanning mean MORE dodging and LESS time spent closing to the 100-200px band where our hit rate peaks (27% vs 5% at 450px). Worth an A/B before drawing conclusions, but it is an experiment, not a shipped change. |
||
|
SirStone
|
69debbe347 |
Re-adaptation: the sliding WINDOW is a control-validated +9.3pp on late accuracy
The user's first-hand diagnosis: "once the pattern is learnt enough we hit DrussGT, but as soon as it adapts we are not fast enough to re-adapt again." The gun kept EVERY sample for the whole battle (which the user explicitly asked for), so stale evidence weighed the same as new evidence - accumulation without forgetting. TASK 1 - THE CURVE, MEASURED FIRST (prequential side accuracy, 15 rounds x 4 horizons, deciles of the eval stream): arm early% late% decay frozen (early only) 64.9 61.4 -3.5 accum (shipped) 76.4 75.3 -1.1 window N=150 84.7 84.6 -0.0 resetdrop 5pp 83.2 84.3 +1.0 **The decay IS real but modest and LOCALIZED IN THE LAST ~30% of the round** (last-third vs middle-third: frozen -7.7pp, accum -3.7, window -1.3, resetdrop -1.3). TASK 3 - WHICH FIX HELPS (late-half accuracy, shuffled control in parens): accum (shipped) 75.3 **window N=150 84.6 (51)** +9.3pp **resetdrop 5pp 84.3 (51)** +9.0pp rehearse-all (retrain, NO forgetting) 79.5 - **Sliding window: +9.3pp late (within-round), +9.7 (cross-round), +10.3 (shield).** The shuffled control stays ~50-51%, so it learns the ENEMY, not noise. - **Forgetting is the essential ingredient**: the same periodic retrain WITHOUT forgetting reaches only ~79.5%, so roughly half the gain is the retraining mechanics and half is the forgetting. - Change detection ties the window on late accuracy and gives the best decay, but at a 5pp threshold it fired 300-500 times in the offline stream - noisy. - INERTIA IS REDUNDANT WITH THE WINDOW: lower inertia helps the keep-everything model a lot (accum late 75.3 -> 83.3 at N=8) but leaves the window FLAT (84.6-84.7 at every N). Low inertia and forgetting are SUBSTITUTES, and forgetting is the robust one - `TR_TMHORIZON_NSTATES=8` is NOT the primary fix. VERDICT: SHIPPING CANDIDATE = `TR_TMHORIZON_WINDOW=150`, kept at default 0 until a live A/B confirms. HONEST CAVEATS THAT SET EXPECTATIONS: - The fixtures are OPEN-LOOP (DrussGT does not react to our bullets), so a true mid-round adaptation is NOT present; the dominant measured effect is the LEVEL gap, not the decay magnitude. The user's "it adapts" magnitude is still INFERRED. - The harness uses a STRAIGHT-LINE base while the live gun uses Pattern's prediction, and it ignores the h-tick label delay, so its absolute side accuracy (75-85%) is INFLATED: the same gun measured ~52% = chance live against DrussGT. So +9.3pp is a real ARM DELTA, not a promise that the gun now clears the ~80% accuracy wall that hits need. A live A/B must decide. Adds `measure_tm_readapt.nim` (prequential harness with the mandatory shuffled control, within-round + cross-round protocols, inertia sweep) and its captured results. test_tm_horizon 79 -> 104 (five new groups). test_tm_diag 48, test_tm_automata_diag 55, test_tm_clause_shape 66, test_rack_membership 48 pass. ModularBot compiles. .gitignore: switched from a broad `measure_*` pattern to EXPLICIT binary names. The broad rule was too blunt - it also excluded the `.txt` results file, which made `git add` refuse the whole commit twice. Sources stay tracked; binaries do not. |
||
|
SirStone
|
b68707c867 |
Energy economy: the cliff becomes a SLOPE, plus a finishing cap. 11% less energy.
The user's request: "when our bot is low OR enemy is low, it is useless to use high power instead low fast bullets have more chances to finish the enemy. Let's do a math slope: starting from some health down, the power goes down with it." 1. ENERGY SLOPE (`TR_POWER_ENERGY_*`), replacing the old hard step at 50 energy: cap = ENERGY_MAX at/above ENERGY_HI, ENERGY_MIN at/below ENERGY_LO, LINEAR in power between, clamped. Defaults HI=80 LO=20 MIN=0.5 MAX=3.0, so no cap >=80, 0.5 at <=20, and e.g. E=65 -> 2.375, E=50 -> 1.75, E=35 -> 1.125. Rationale: bullet speed is 20-3p, so lower power = FASTER bullet (less lead error, higher hit chance), fires more often (10+2p) and drains slower (p/shot). E[dE] = p(3P-1) => break-even hit probability is 1/3 INDEPENDENT of power, and our measured rates are 5-27%, far below it. 2. FINISHING CAP (`TR_POWER_FINISH_KILL`, default ON): cap power at the SMALLEST bullet that still removes the enemy's remaining energy - `E<=4 -> p=E/4` (min 0.1), `4<E<=16 -> p=(E+2)/6`, `E>16 -> no cap`. Rationale, and it makes the user's instinct stronger than a heuristic: server 1.3.1 caps the damage SCORE at the energy ACTUALLY REMOVED, so overkill is WASTED damage AND ~6x the energy for ZERO extra score. Damage is 4p (p<=1) / 6p-2 (p>1). Both are min-composed with the existing far/below-average caps, may only LOWER power (exhaustively tested), and are exempt while ramming. `TR_POWER_POLICY=0` still returns the uncapped control exactly. MEASURED ENERGY SAVING (offline replay of the DrussGT fixtures, 28,797 ticks): arm shots energy meanP E/1k ticks vs cliff control(uncapped) 1913 4646 2.43 161.4 -90.2% cliff (today) 2363 2443 1.03 84.8 0.0% slope 2404 2178 0.91 75.6 ** 10.9% LESS ** slope+finish 2404 2167 0.90 75.3 ** 11.3% LESS ** So the slope spends ~11% less energy than the cliff AND fires slightly MORE shots (2404 vs 2363) - both directions at once. HONEST NOTE on the finishing rule's reach here: ticks where the enemy is low (0 < E <= 16) are only 2252/28797 = 7.8% of these fixtures, so finishing adds just ~11 energy of saving against DrussGT. It matters in CLOSER fights, not this one. Verification: test_power_policy 58 (was 26) in BOTH the default and TR_POWER_POLICY=0 control arms - slope at E=100/80/65/50/35/20/5, powerToKill across E=0.1..100, the inverse-cover property for E<=16, monotonicity, ram exemption, and an exhaustive sweep proving power <= preference. Guards: test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_ram_decision 40, test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20, test_vbullet_admit_gate 12. acceptance 12/12 PASS. ModularBot compiles release. Adds `common_libs/tests/measure_power_policy.nim` (the energy/histogram tool) and updates docs/env_reference.md for the new `energySlope|finishKill` log reasons. NOT MEASURED: the battle/hit-rate effect. The offline figures use the fixture shooter's energy as a proxy, open-loop; the RELATIVE saving is the meaningful part. |
||
|
SirStone
|
81af5854df |
FIX an incomplete commit: TMHORIZON was admitted unconditionally at HEAD
MY ERROR. `aed579b` committed `ModularBot.nim` (which wires the new gun as id 15) and `guns/tm_horizon.nim`, but I staged only three files and left job-72's RACK REGISTRATION behind in the working tree. Consequence at HEAD: - `common_libs/gun_harness/selector.nim` still had the 15-entry rack table (ids 0..14) with no TMHORIZON name, so - `TR_RACK_TMHORIZON` was never read (a silent no-op), and - id 15 is PAST the table, and `rackAdmitted` documents "an id past the membership table is admitted" -> **TMHORIZON was admitted unconditionally**. So the committed default rack was `Pattern + TMHORIZON`, not the intended `Pattern only`, and the gun could not be switched off by env at HEAD. This was caught by a live A/B job that had to work around it with `GUN_RACK_DISABLE=15`. FIX (this commit): the rack table goes to 16 entries with `TMHORIZON` appended at id 15 defaulting to `rmOff`, plus the two tests whose length literals follow it (`test_rack_membership` 15->16, `test_tm_pattern_registration` 15->16). `DefaultRackMembership` is again Pattern-only with every other gun `off`. NOTE ON SCOPE: `selector.nim` in the working tree also contains a CONCURRENT job's power-policy threading (an `enemyEnergy` parameter). That work is still in flight and is deliberately NOT included here - only the three rack hunks were staged. The remainder stays unstaged for its own commit. The lesson, recorded because it has bitten twice tonight in different forms: a change is not committed until its registration/table counterpart is, and `git add` of a hand-picked file list is exactly how a half-change ships. |
||
|
SirStone
|
9ba932d1b1 |
docs: complete environment variable reference (every knob, its default, and why)
Single place to answer 'what env vars exist, what do they default to, and what are
they for'. Grouped by area, with the measured reason each knob exists recorded
next to it, since several defaults are counterintuitive:
- The shipped rack is PATTERN ONLY, with the revert one-liner included.
- GUN_SELECTOR_TIEBREAK defaults off because it measured NEGATIVE on real hit
rate, and the per-tick random draw inside the tie band is load-bearing
(commitment cost 7.02% -> 5.10%, p=0.002).
- The power policy is ON because turning it off measured 10.6% -> 7.9% real hit
rate; TR_POWER_FINISH_KILL exists because server 1.3.1 caps the damage SCORE at
the energy actually removed, so overkill scores nothing.
- The ring mover is NOT the default and must not be shipped: best offline hit
rate of anything measured, but it halves survival (16/49 -> 6/49, p=0.012).
- The proactive ram is off because it converts 0/6 times.
- TR_PATTERN_RAD_* are kept but are structurally incapable of changing the shot
(the aim is bearing-only, measured byte-identical on the path metric).
- TR_TMHORIZON_NSTATES is the automata inertia ('mood'): lower = adapts faster.
- TR_VBULLET_ADMIT_ONLY=1 gave +68% tick rate (87 -> 146 ticks/s).
Documents the read-once-at-start convention and the operational trap that the bot
is spawned by the server/GUI, so a variable exported in an unrelated terminal does
NOT reach it. Also lists the compile-time -d: knobs and the test-harness jars.
Snapshot of the source at this commit; f9f8d84-era knobs added by the in-flight
jobs are included where already present.
|
||
|
SirStone
|
aed579b3af |
TM horizon: retain learning ACROSS ROUNDS, reset only when the ENEMY changes
The user's requirement: "every battle i means from round 1 to round end-battle, so retain all learning until the enemy change." What was built wiped the Tsetlin machines EVERY ROUND, in two places (`onRoundStarted` and the gun's own tick-regression self-reset), so in a 7-round battle each round started cold, trained ~360 samples and threw them away - discarding most of its one chance to do what was asked: overfit the current enemy over the whole battle. THE FIX - two kinds of state, two triggers: - **`resetRoundState` (per ROUND)**: the observation ring, pending/deferred labels, the bullet proxy, motion history, per-tick caches, `roundStartTrained`. These MUST clear every round, because bots teleport back to the starting corners between rounds - an old position would build a garbage label. (That exact class of bug shipped 36-58% wrong labels in the old gun.) - **`resetLearning` (per BATTLE / per ENEMY)**: both Tsetlin machines, `trained`, `sideCorrect/sideTotal`, all histograms, the magnitude median, `pendingDropped`, `observedTargetId`. These now SURVIVE round boundaries. Triggers for the machine wipe: `onGameStarted` (primary) plus a redundant `roundNumber <= 1` fallback in `onRoundStarted`; and a TARGET CHANGE (`targetChanged`, knob `TR_TMHORIZON_RESET_ON_TARGET` default on - a no-op in 1v1, fires on melee target switches; first acquisition never wipes). The tick-regression self-reset now clears ONLY per-round state. Still NO cross-battle persistence: grep for file I/O in the gun finds none. PROOF IT WORKS (live 2-round battle, `TR_RACK_PATTERN=off TR_RACK_TMHORIZON=both`): [tmh-reset] reason=game_start trained_was=0 [tmh-reset] reason=round1 trained_was=0 ...exactly TWO reset lines in the whole battle, both at battle start, and NONE at the round-2 boundary. And the per-round summaries: [tmh-round] trained=1249 thisRound=1249 ... sideAcc=711/1177 (60.4%) [tmh-round] trained=2226 thisRound=977 ... sideAcc=1361/2154 (63.2%) `trained` CLIMBED 1249 -> 2226 across the boundary, and side accuracy rose 60.4% -> 63.2% in round 2 (one battle - suggestive, not proof). Unit tests: `test_tm_horizon` 79 (was 54), including "trained SURVIVES the boundary", "clause states SURVIVE", "trained climbs round1->round2", "game-start wipes and all clauses end Exclude", "different enemy wipes / same enemy does not / knob-off does not", and crucially "a label CANNOT be built across a round boundary" (ringValidCount==0, ringHas(oldTick)==false, pendingCount==0) - the single most dangerous interaction of this change. Guards: test_tm_horizon 79, test_rack_membership 48, test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 40, test_selector_tiebreak 19, test_tm_pattern_registration 20, test_vbullet_admit_gate 12, test_tm_diag 48, test_tm_automata_diag 55, test_tm_clause_shape 66. acceptance_offline_vs_online 12/12 VERDICT PASS. Shipped rack unchanged: DefaultRackMembership is still Pattern-only, TMHORIZON off. Residual (pre-existing, out of scope, stated): the internal base `PatternMatcherGun` has a rolling move-history buffer that is NOT cleared at round boundaries - it never was, and the SHIPPED Pattern gun carries history across rounds too. It cannot affect label correctness (labels come from `g.ring`), only base-prediction quality in a round's first ticks. |
||
|
SirStone
|
be74e369eb |
Gate 2b: the shift form WORKS - but break-even needs ~80% side accuracy and the
problem gives 60%. Thread closed with a number, not a shrug.
Gate 2 (
|
||
|
SirStone
|
fd2f7f608c |
GATE 2: STOP - the TM learns the side but the correction cannot buy hits
Offline four-arm pipeline: extracts the 49-bit draft spec + a 4-bit horizon one-hot (53 bits) from the DrussGT fixtures, derives fact-based quadrant labels (side vs the naive guess, magnitude vs the TRAIN median), trains the validated `tm_diag/tm_core.nim` fresh PER ROUND with a fit / calibrate / eval within-round split, applies each predicted quadrant's out-of-sample conditional-median offset and measures the residual angular error. 50 clauses, N=64, s=3.0, 5 epochs (verdict identical at 1 and 10). Estimated hit = |residual| < atan(18px/range). POOLED (tr_drussgt_vs_modularbot + _shield, N~9.1-9.4k per horizon): h arm med|err| p90|err| hit% 15 naive 4.08 14.24 39.3 15 TM 4.81 13.53 29.0 15 shuffled 4.45 14.58 34.3 15 turn-only 4.13 14.10 37.1 20 naive 6.81 21.60 28.2 20 TM 7.61 20.85 19.2 25 naive 9.71 29.13 20.9 25 TM 10.33 28.63 13.1 30 naive 12.62 36.12 16.4 30 TM 13.09 35.17 10.5 Delta hits (pp): TM-naive **-10.3 / -9.0 / -7.8 / -5.9**; **TM-turn -8.1 / -5.1 / -2.0 / -0.9** (h=15/20/25/30). Median miss: TM is WORSE by +0.5..+0.8 deg. p90: marginally better by 0.5-1.3 deg. So it pulls in the TAIL but not the typical miss. Shuffled control: quadrant accuracy 22.9-24.8% ~= 25% chance, and dHit <= 0 at every horizon -> NO LEAK. (The mild negative is the expected cost of applying a noisy offset, not leakage.) THE MECHANISM, and it is the important part: **the learning is REAL - the TM gets the side right 59.6-61.0% vs 50% (quadrant 34-35% vs 25% chance; turn-only 53-56%) - but it does not translate into hits because the per-quadrant conditional-median offsets (~4-16 deg for "big") are FAR LARGER than the body half-angle (~2.6-3.4 deg at typical range).** The naive guess is essentially UNBIASED (the median signed error is 0.00 deg at every horizon, from the headroom study), so its error is centred on the target; shifting an already-centred distribution away from zero DESTROYS near-target mass. That is why the turn-only arm loses too (-2.2 to -5.8 pp): any constant shift hurts. VERDICT AS IMPLEMENTED: **STOP. There is no case for building the new TM gun on this evidence.** NOT YET DISTINGUISHED, and worth one cheap test before the idea is declared dead: whether this is a STRUCTURALLY dead application (no shift can help, because the baseline is unbiased and the correction is coarser than the target) or merely a MISCALIBRATED one (the offset was fitted to minimise the conditional MEDIAN of the error, which is NOT the objective - hits are maximised by the shift that maximises P(|error| < body), typically a SMALLER shift or none at all on a dense near-zero distribution). That distinction decides whether the whole "TM predicts the enemy's position" premise is dead or only this instantiation of it. Caveats: DrussGT-only; the enemy's movement is a closed-loop response to our CURRENT movement so the numbers are conditional on how we move now; offline observation is perfect every tick while live we see the enemy only on scans, so all of this is an UPPER BOUND. The bullet block was proxy-based (no gun heading or power is recorded in the fixtures) - INFERRED, and stated as such rather than silently dropped. |
||
|
SirStone
|
9064377740 |
s settled from OUR code (higher = LONGER clauses), and it cannot rescue the gun
=== TASK 1: THE DIRECTION QUESTION, ANSWERED WITH A DEMONSTRATION ===
I told the user `s` controls clause length but refused to claim the DIRECTION,
because I had seen it described both ways. It is now read out of our own code -
one site per core, in the Type I branch of `tmLearnDir` (`guns/tm_pattern.nim:258`,
`guns/tsetlin.nim:224`, `tm_diag/tm_core.nim:110`):
if pol * d > 0.0:
if lits[lit] == 1:
if cOut == 1: if rand < (s-1)/s: st += 1 # toward Include, w.p. (s-1)/s
else: if rand < 1/s: st -= 1 # toward Exclude, w.p. 1/s
else: if rand < 1/s: st -= 1 # toward Exclude, w.p. 1/s
=> **HIGHER `s` GIVES LONGER CLAUSES.** The include step runs w.p. (s-1)/s
(rising with s); both exclude steps run w.p. 1/s (falling with s).
DEMONSTRATED (49-bit draft, planted 2-literal rule, 3000 train / 1500 eval):
s=1.0 len 1.51 acc 100% s=2.0 len 1.82 acc 100% s=5.0 len 3.02 acc 100%
s=1.5 len 1.42 acc 100% s=3.0 len 2.19 acc 100% s=10 len 4.04 acc 99.7%
s=20 len 5.17 acc 94.5%
WHY s=1.0 DEGENERATES: (s-1)/s = 0 so the include step NEVER fires while 1/s = 1
so BOTH exclude steps always fire - Type I can only remove literals, so a clause
can grow only through the Type II penalty. (On random labels that leaves 43/120
non-empty clauses vs 120/120 at s>=3.)
USABLE RANGE ~[1.5, 5]. tm_pattern uses 3.0; tsetlin uses 1.5.
=== TASK 4: WOULD `s` HELP THE SHIPPED GUN? NO - MEASURED ===
Recompiling the offline driver with -d:TM_S_DEF=<v> (source untouched) retrains
the gun end to end:
s mean len verdict warm acc margin vs majority
1.5 15.82 too long 32.22% -2.03pp
2.0 14.99 too long 34.03% -0.22pp
3.0* 19.17 too long 35.72% +1.48pp (*shipped)
5.0 20.47 too long 34.72% +0.47pp
Lowering `s` shrinks the clauses and makes accuracy WORSE; raising it pads them
and also loses. The shipped 3.0 is the best of the four, and **no value comes
near the healthy 3-8 band.** Combined with the settledness finding, the shape is
consistent with "NO CONSISTENT SHORT RULE EXISTS in this representation/target".
So the bottleneck is the SIGNAL - now confirmed from a THIRD independent angle
(settledness, churn trend, and clause shape). This is the measurement behind the
decision not to spend effort sweeping N or s.
=== TASK 3: AN HONEST CORRECTION TO MY OWN HYPOTHESIS ===
I predicted that random labels would produce `too long` clauses (the TM padding).
MEASURED: on this encoding noise reads as **short / `collapsed`** (mean 1.88,
median 2.0, acc 33.3%) - the TM FAILS TO COMMIT rather than padding. So "too
long" is not the noise signature, which means the shipped gun's 19.17 mean is not
explained by label noise. Worth knowing.
Adds diagnostic group 8: the clause-shape checker - full length distribution
(min/median/p10/p90/std), per-polarity and per-class breakdowns, a
`clauseShapeVerdict` against a parameterised healthy band (default 3-8),
per-BLOCK length contributions, and clause coverage (mean firing clauses,
effectiveClauses = participation ratio, top3Share). `healthLine` now appends
`shape=<mean> (<verdict>)`.
Validation: `test_tm_clause_shape` 66 checks. A planted 2-literal rule reads
`healthy` with the literals recovered exactly; per-block correctly names the
planted blocks (WALLS 43.0%, BULLETS 28.1%) and buries an irrelevant block (3.4%,
below its uniform 8.3% share); random labels read `collapsed`.
REAL READING, shipped gun: mean 19.17 / median 16.00 / p90 44.80 / max 57,
173 non-empty of 200, 27 empty => **`too long`**; coverage firing/sample 48.53
(24.3%), effectiveClauses 97.85/200, top3Share 5.3% (voting NOT concentrated);
per-block is diffuse with no dominator, EXCEPT **UNUSED 6.4%** - the always-true
negations of the never-written bits 38/39 acting as FREE PADDING, the same bug the
kit found earlier now visible as clause bloat.
Guards: test_tm_clause_shape 66 (new), test_tm_diag 48, test_tm_automata_diag 55,
diag_synthetic 17, diag_automata_validation 11, test_gun_harness 39,
test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41,
test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 40,
test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20,
test_vbullet_admit_gate 12. acceptance_offline_vs_online not run (needs a live
battle; no tm_diag dependency).
|
||
|
SirStone
|
e180626b50 |
Horizon headroom: QUALIFIED PASS at h=10..50, but the signal is one feature
Measures the learnable headroom of the "where will the enemy be in h ticks" problem on the DrussGT fixtures, for h=1..50, using the ANGULAR error (the aim only cares about the angle - a distance-only error cannot change the shot). Observer = ModularBot at t; naive guess = straight-line extrapolation, never bounced off walls. Round-bounded via the .rounds.json sidecars; the tail h ticks of each round are dropped, never labelled with garbage. N = 32,630 (h=1) down to 31,405 (h=50). WHAT THE NUMBERS SAY - **The question is FAIR.** Sign balance is dead-centre at EVERY horizon (47.7-50.2% left) with zero systematic bias (median signed error = 0.00 deg at every h). No de-biasing needed - a healthy symmetric question, which is what the two-binary output shape needs. - **The dumb guess is exact short, wrong long.** naiveMiss (aim lands outside the 18px body): 0% at h<=3, 3.3% at 4, 11.8% at 5, 27.6% at 6, 48.6% at 10, 62.6% at 15, 73.5% at 20, 85.9% at 30, 93.7% at 50. Error magnitudes: median 2.0 deg (h10), 7.0 (h20), 12.8 (h30), 18.5 (h40), 24.5 (h50). Body half-angle at 300px is 3.43 deg - so at long horizons the error is 4-7x the body size. - **No trivial rule solves it.** Turn-direction accuracy is 53-54% at h<=3, drops through 50% at h~7, and INVERTS to 38-45% at long h (flipped = 55-62%, the best single rule anywhere). Causal 1-step persistence peaks ~65% at h2-4 and decays to ~49% by h50. Majority is 50-51.5%. - **REAL signal exists in exactly ONE feature: the enemy's current turn/reversal direction.** dTurn swings from +8.4pp (h2) through 0 (h7) to **-24.0pp at h50", all |z|>6. Every other planned feature is WEAK: walls <=2-4pp, speed <=5pp, bullets <=4pp, reversal <=6pp, closing <=3pp. TWO FINDINGS I DID NOT EXPECT 1. **The dead zone is h=5..9.** The naive guess starts missing there (12-44%) but NO tested feature shifts the left/right split by >=5pp - the sign is a featureless coin flip in that band. So those horizons carry no learnable signal despite looking promising. 2. **The bullet block - which we designed with enthusiasm - shows <=4pp of shift.** The enemy's dodging reaction to our bullet is NOT a strong conditioning signal at these horizons in this measurement. CAVEAT: the "bullet in flight" split used an energy-drop proxy with ~370 false positives in 1504 positives, so this is a weak negative, not a settled one - the block should be measured properly before being cut. RECOMMENDED RANGE: **h = 10..50** (41 horizons, contiguous). Criteria: (a) 42-58% left, (b) max(turnAcc, pers1Acc) < 75%, (c) some feature |delta| >= 5pp with |z| >= 3, (d) naiveMiss >= 15%. h=1-4 fail (d); h=5-9 fail (c); from h=10 all four hold and strengthen with h. VERDICT - QUALIFIED PASS, and the job's own calibration is worth quoting: there is genuine, non-degenerate structure, so the horizon-input design is not obviously wasted; BUT the per-sample signal is weak and concentrated almost entirely in one feature which already captures most of the easy structure (~60% sign accuracy). **"I would not treat this as a green light for a big build; I would first check that a model can beat 60% sign accuracy on a held-out round at h~15-25."** CAVEATS (from the tool): DrussGT-only, and the enemy's movement at capture time was a RESPONSE to our current movement, so the headroom is conditional on how we move now; offline observation is perfect every tick while live we see the enemy only on radar scans, so these numbers are an UPPER BOUND; and the replay is open-loop even though the capture was closed-loop. |
||
|
SirStone
|
ab8d383121 |
Automata metrics: settledness alone does NOT separate learning from fidgeting
Added the four automata-level metrics to the TM diagnostics kit (settledness, clause diversity, churn, vote disagreement) plus a state histogram, a per-input confidence table and a one-line health summary, and validated them on a learnable-vs-noise pair. STATE CONVENTIONS, read off OUR code rather than from memory: range [-nStates, nStates] as int16; nStates = 64 for tm_pattern, 32 for tsetlin initial value 0 = the Exclude boundary INCLUDE iff state > 0; EXCLUDE iff state <= 0 flip boundary sits between state 0 and 1; commitment = abs(st)/nStates in [0,1] === THE GATE, AND A RESULT THAT MATTERS === Case A (learnable planted rule) vs Case B (shuffled labels), 49 bits, N=64: metric A (learnable) B (shuffled) settledness mean 0.970 0.719 churn flip/sample 0.000055 FALLING 0.000788 FLAT clause-change/sample 0.00263 falling 0.0595 flat diversity (Jaccard) 0.176 0.014 disagreement 0.003 0.298 verdict settling mixed (NOT settling) **SETTLEDNESS ALONE DOES NOT WORK.** On noise the automata still COMMIT (0.719) - they just commit to the wrong thing. The decisive separators are **churn TREND (falling vs flat)** and **vote DISAGREEMENT (0.003 vs 0.298)**. Had we built only the settledness metric - the one that seems most obvious - we would have been misled. That is now recorded in the README. INERTIA SWEEP: A vs B separate at N=16/32/64/128. **Raising N raises A's commitment but does NOT reduce B's noise-fitting** - so more inertia does not rescue a noise-fitting TM. === REAL READING ON THE SHIPPED GUN, AND THE INFERENCE IT SUPPORTS === tm_pattern GF head over the DrussGT fixtures: settledness 0.484 (settling), diversity 0.267 (moderate), churn 0.094/100 FALLING, disagreement 0.145 (coherent). **VERDICT: SETTLING** - not fidgeting, not collapsed. Constant inputs flagged: 38/39 (the known never-written bits) plus 19/36/37. Context: pooled warm accuracy 35.72% vs 34.24% majority = +1.48pp. So: **the old gun was NOT failing because of inertia or instability - it settled properly and its settled rules still barely beat a lazy guess.** Its settledness (0.484) is LOWER than both synthetic cases (0.97/0.72), which is the signature of WEAK OR CONFLICTING SIGNAL rather than too much inertia. CONCLUSION: **N and s are not the observed bottleneck. The target/representation is.** That is exactly why the new design changes the target and the label pipeline rather than sweeping knobs - and it means we should NOT spend effort on an N/s sweep expecting it to fix anything. Also adds `diag_automata_validation.nim` (Case A/B/C + inertia sweep) and `test_tm_automata_diag.nim` (55 pure checks); `test_tm_diag` 48 and `diag_synthetic` 17 still pass, plus all other guards. acceptance_offline_vs_online was NOT run (it needs a live battle and there is no tm_diag dependency). Caveat: churn on the real gun is a PROXY (a tm_core retrain over captured samples in live order) because the live gun exposes no per-sample state trace; the other metrics are read directly off the exported teams. |
||
|
SirStone
|
f9f8d84671 |
TM diagnostics kit: VALIDATED (finds a known dead input), and it found a real bug
Built `common_libs/tm_diag/` as a first-class offline diagnostics kit for Tsetlin
work, BEFORE writing the new gun - because we hit two data problems tonight that no
amount of reading the TM's clauses would have revealed (a 38.8% majority answer,
and 36-58% mislabelled training samples).
WHAT IT PROVIDES
- `feature_spec.nim`: a NAMED feature container, so a learned clause prints as a
sentence (`IF near-wall AND bullet-dead-on AND turn-left(t-2) THEN class=3`)
instead of "feature 17". Includes the 49-bit draft spec from the design session
and the shipped 40-bit encoding.
- `tm_core.nim`: a compact deterministic Granmo multiclass TM with an
INTROSPECTABLE clause layout (mirrors the tm_pattern core).
- `diagnostics.nim`, six groups: (1) pre-flight DATA checks + shuffled-label
control, (2) clause introspection (readable dump, per-clause vote counts, empty
and never-fired clauses, length distribution, per-class balance), (3)
per-feature contribution with an explicit DEAD-INPUT LIST and a ranked
most-valuable list, (4) accuracy vs the majority baseline with per-class
precision/recall and pred-majority share, (5) learning curve, (6) ablation hooks
(drop a block / scramble a bit).
=== TASK 3: THE VALIDATION THAT GATES EVERYTHING - PASSED WITH NUMBERS ===
A diagnostic we never checked is worthless, so the kit was tested on a synthetic
set with a PLANTED RULE (class2 = A and B, class1 = A and not B, class0 = not A),
a deliberately IRRELEVANT block (US, 9 bits) and a PURE-NOISE bit (17).
- majority baseline 60.63% (class0); over-30% correctly flagged
- **the planted rule is recovered EXACTLY** via `necessaryLiterals`:
class0 IF NOT dist-wall<50 | class1 IF dist-wall<50 AND NOT lat DEAD-ON |
class2 IF dist-wall<50 AND lat DEAD-ON
- **DEAD-INPUT LIST = all 9 US bits AND the noise bit 17**, while the planted bits
0 and 45 are correctly NOT listed
- top contributors: bit0 w=1241.7, bit45 w=583.3, then 49.8 - a 12-25x gap, so the
relevant bits are unmistakable
- **ABLATION: drop WALLS -39.47pp, drop BULLETS -19.33pp, drop US 0.00pp**,
scramble A -42.00pp, scramble the noise bit 0.00pp
- shuffled-label control 60.40% vs majority 60.63% = -0.23pp -> no leak
So the kit reliably finds a known dead input and a known relevant one.
=== TASK 4: THE REAL READING, AND A BUG IN THE SHIPPED GUN ===
`tm_pattern` GF head, 6 DrussGT fixtures, pooled 250,745 samples:
- label balance c2 = **34.4%** (majority-heavy, flagged); accuracy **35.72%** vs
majority **34.24%** -> margin **+1.48pp**. On `tr_drussgt_vs_crazy` it is BELOW
majority (33.81% vs 37.72%, -3.92pp).
- 200 clauses: **27 empty, 45 never fired**, mean length 19.17, max 57. The
majority class is starved (class2: 22 non-empty, 18 empty, only 2 positive
fired). Class4 fires 11-24-literal clauses -> memorisation signature.
- **REPRESENTATION BUG FOUND (reported, not silently fixed):** `tmBuildBits` writes
only 38 raw bits into `var bits: array[TM_NBITS=40, uint8]` - bits 38 and 39 are
NEVER ASSIGNED, so they are always 0 and their negated literals are always 1.
The kit's `constantInputs` confirms 38/39 are constant, and **`UNUSED-38`/`39`
rank #6 and #8 in the most-valuable-inputs list** - i.e. the model's
highest-usage inputs are information-free. That is a representation bug, not a
display artefact, and it is a concrete mechanism for part of the poor learning.
DRAFT ENCODING CHECKED: the 49-bit draft is arithmetically consistent
(4+4=8 walls, 6+3=9 us, 3+5+3+3+3+3=20 motion, 5+7=12 bullets = 49). No draft
inconsistency.
Guards: test_tm_diag 48 (new), diag_synthetic 17 (new), test_gun_harness 39,
test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41,
test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 40,
test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20,
test_vbullet_admit_gate 12, acceptance_offline_vs_online 12/12; tm_pattern_learning
passes. The tm_pattern hook is additive and default-OFF (no behaviour change).
NOT YET INCLUDED (the automata metrics discussed for the next step): per-clause
automata settledness (distance from the flip point), clause diversity (pairwise
overlap), literal-set churn over time, and cross-clause vote disagreement. The kit
has clause-level diagnostics but not the automata-state ones.
|
||
|
SirStone
|
b0654d18eb |
TM verdict, settled: it loses LIVE and sits at/below its majority class - (c)
The user pushed back on "the TM can't be your best 1v1 gun", correctly, because two
decisive tests had never been run. Both are now run and they agree.
TASK 1 - THE GF HEAD vs ITS MAJORITY-CLASS BASELINE (offline, n=1,751,067):
label histogram [254286, 284578, 678879, 297055, 236269]
majority class = 2 (the CENTRE bucket) = 38.77%
RAW head accuracy = 36.69% -> margin **-2.08 pp, BELOW majority**
GATED head accuracy = 40.37% vs 38.75% majority -> +1.62 pp, BUT it predicts the
majority class on 62.4% of ticks and its minority recall is 13.6% / 12.9% - a
base-rate predictor wearing a classifier's clothes.
Shuffled control sits at its own majority (20.04% vs 20.12%), confirming chance.
**THE OLD "46% vs 20% CHANCE" FIGURE I QUOTED WAS WRONG ON TWO COUNTS:** the
baseline is 38.8%, not 20%, and the 46% predated the deferred-label fix. Against
the correct baseline the head is BELOW it.
TASK 2 - THE FIRST-EVER LIVE A/B OF THE TM GUN (7 runs x 7 rounds per arm, one
frozen binary from git archive HEAD =
|
||
|
SirStone
|
eb74f9b2e3 |
Ram: finisher-only by default, and the bullet-rain abort now measures real energy
Follows the diagnosis that proactive straight-line ramming CANNOT work: both bots have MAX_SPEED=8, so a pursuit cannot catch an evading equal-speed opponent. Measured over 49 rounds per arm, opportunity -> contact was **0/6** (base), 0/40 (ring), 0/12 (ringhot). The only proactive conversion in the whole corpus came from a FINISHER, and only because a <20-energy DrussGT stops fleeing (that episode closed at 6-8 px/tick). Opportunity episodes never got below ~80px; one ran the full 60-tick duration cap and closed only 198->171px; a perfectly aligned full-speed one closed 195->114px then plateaued. CHANGES - **Finisher-only default.** `finisher` (<20 energy, dist<300, we are healthier) and the rare `desperation` (both <5, dist<150) are kept; `opportunity` and the speculative `plan` are OFF. Both are env-reenableable with no rebuild: `TR_RAM_OPPORTUNITY=1` (tune via TR_RAM_OPP_DIST/MARGIN) and `TR_RAM_PLAN=1`. Justification: it removes 100+ non-converting episodes per fixture at zero measured loss (oldram vs base was p=0.69, damage 279 vs 284, survival 17/49 vs 16/49) - and each of those episodes spent up to 60 ticks driving STRAIGHT at the enemy, abandoning the mover's dodging and disrupting aim. - **`desperation` KEPT** deliberately: it is cheap and rare, fires only when both bots are nearly dead at short range (a coin-flip where 0.6 contact can decide it), and it is not the refuted straight-line pursuit. - **THE BULLET-RAIN ABORT WAS DEAD CODE AND IS NOW FIXED.** `onHitByBullet` accumulated raw bullet FIREPOWER while `TR_RAM_ABORT_DMG = 0.5` was documented as a DAMAGE rate - so the bar was implicitly "sum of power > 7.5 over 15 turns" and the maximum rate ever observed was 0.27. It now accumulates REAL ENERGY via a `bulletDamage(power)` helper matching the server's `4p` / `6p-2` formula, and `TR_RAM_ABORT_DMG` defaults to **2.0 energy/turn** (~30 HP over 15 turns): "abort an in-progress ram if we take > 2.0 energy per turn". Same effective bar for normal firepower, and it can now actually fire - the live run reports `dmgRate=1.07/turn` where the old units said 0.27. - **`ramStuckTicks` REMOVED.** It required `dist < 5px`; contact occurs at ~36px (two 18px radii) and position rewind prevents getting closer, so it could never increment. Only the 60-tick duration cap can now self-end a ram. LIVENESS (measured, default config, vs a charging Java RamFire, 3 rounds): default -> `[ram] ON reason=finisher` x3, `reason=opportunity` x0 TR_RAM_OPPORTUNITY=1 -> `reason=opportunity` x4, `reason=finisher` x2 So the opportunity states DID occur and are suppressed by the new default - the removal is real, not an arm that never fires. A line also read `[ram] OFF reason=duration dmgRate=1.07/turn`, confirming the new energy units. Adds docs/ramming_negative_result.md (70 lines) recording the question, the five diagnostic answers, the geometric reason, the finisher exception, the two dead code paths, and an explicit "do not re-attempt a proactive straight-line ram; if point-blank forcing is ever wanted it is an INTERCEPTION/cornering movement problem" note - the same pattern that stopped the corpse bug recurring. Guards: test_ram_decision 40 (was 28), test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_rack_membership 48, test_selector_tiebreak 19, test_tm_pattern_registration 20, test_vbullet_admit_gate 12, acceptance_offline_vs_online 12/12. ModularBot compiles. Honest note: the abort-threshold fix is a real (tiny) behaviour change, NOT measured-neutral - it only bites while a finisher ram is under sustained fire, which is exactly the user's stated wish. The finisher-only removal itself is measured-neutral per the given A/B. |
||
|
SirStone
|
99f9532f55 |
Melee vs 1v1 racks: mechanism built, but NO gun-level payoff - do not split
The user's plan was separate melee and 1v1 racks. The mechanism is built and
committed (TR_RACK_<GUN>=both|1v1|melee|off, mode from server truth). This job
produced the missing evidence: each candidate gun forced ALONE in BOTH modes,
one frozen binary, env-only arms, exact two-sided permutation tests.
1v1 vs the real DrussGT (6 arms x 5 runs x 7 rounds):
Pattern 10.80% 292 dmg/run 11/35 round wins
KNN 5.58% 128 1/35
Linear 2.99% 59 0/35
Circular 2.89% 60 0/35
WallBounce 2.72% 57 0/35
GuessFactor 2.14% 38 0/35
-> every arm differs from Pattern at p=0.0079 (the 5v5 floor). Per-gun damage
span 7.7x. In 1v1 the gun matters ENORMOUSLY.
Melee (+DrussGT, RandomMover, WaveSurfer, OscillatorBot; enemyCount 4):
WallBounce 25.79% 643 dmg/run 7/25
Linear 24.88% 624 7/25
GuessFactor 24.48% 613 4/25
Pattern 24.39% 599 5/25
Circular 25.88% 596 5/25
KNN 23.87% 539 3/25
-> NO arm beats Pattern with significance (p=0.42-0.96, fully overlapping).
WallBounce's nominal +7.3% is p=0.42. Per-gun damage span only 1.19x.
THE FINDING: in 1v1 the gun matters enormously (7.7x spread); in melee it barely
matters (1.19x). Melee is won by movement, survival and placement, not by which
gun you carry - every candidate lands in the same ~24-26% band. So a separate
melee rack has NO gun-level payoff and DefaultRackMembership stays Pattern-only.
The mechanism remains available if it is ever wanted.
Honest caveats: the melee comparison is UNDER-POWERED at 5 runs (detecting the
~44-damage WallBounce-Pattern gap would need ~2.5-3x the runs), so "no significant
difference" is NOT "no difference"; the only candidate worth re-testing is
WallBounce in melee, and it must not be shipped on this evidence.
SHIM LIMITATION FOUND, worth recording: run_bridge_battle.sh captures with
`--subject DrussGT`, and the shim DROPS every tick once the subject dies - losing
5-20% of ModularBot's shots in melee. The campaign was re-run with
`--subject ModularBot` so ModularBot's events are complete (fires match gun_stats
realShots exactly). Any future melee evidence through this shim must do the same
or it will silently under-count.
Adds docs/melee_vs_1v1_racks.md. No repository source changed.
|
||
|
SirStone
|
bfdcdf8919 |
The three unmeasured features, A/B'd - and a methodology correction I had wrong
Five arms x 7 runs x 7 rounds (35 real-DrusGT battles, 8 concurrent), one frozen
binary from git archive HEAD at
|
||
|
SirStone
|
3142b70aa5 |
Gate virtual-bullet spawn on rack admission: +68% tick rate, selected gun unchanged
The default rack is now Pattern-only (
|
||
|
SirStone
|
185a32e9eb |
Radial offset: STRUCTURALLY incapable of helping, and Pattern does not overshoot
Follow-up to
|
||
|
SirStone
|
31c7c01d28 |
SHIPPED: the default rack is now Pattern-only (+49% hit rate, +66% damage on the boss)
`DefaultRackMembership` now admits Pattern (id 5) and marks all 14 other guns `rmOff`. **The selector mechanism is untouched** - `chooseFromFit`, the floor/band logic, the hysteresis and the virtual-fitness plumbing are all intact and functional. Only the rack membership changed, so this is reverted by env alone. Evidence (measured, replicated three times, 10 adversaries): Pattern alone gives 10.36% real hit rate / 264 damage per run vs the full rack's 6.93% / 159. Pattern significantly wins on DrussGT, Corners, Crazy and PatternMover, ties on three, and the full rack never significantly beats it on ANY adversary. Mechanism: the virtual signal keeps ranking the wrong guns first (HeadOn 46% of ticks at 2.0% real; Linear 57.7% at 6.0% real while Pattern sits at 11.1%). **THIS CONTRADICTS THE USER'S STANDING DIRECTIVE** to keep virtual-fitness selection. Recorded plainly in docs/selector_negative_value.md with a SHIPPED DECISION banner rather than done quietly: the mechanism is retained and one env var away, because the measurement says it is negative value on every rack size tested and on 10/10 adversaries. Revert one-liner (no rebuild): TR_RACK_PATTERN=both TR_RACK_HEADON=both TR_RACK_LINEAR=both TR_RACK_TSETLIN=both \ TR_RACK_CIRCULAR=both TR_RACK_GUESSFACTOR=both TR_RACK_WALLBOUNCE=both \ TR_RACK_ACCEL=both TR_RACK_STOPSHOT=both TR_RACK_DISPLACE=both TR_RACK_AVGLEAD=both \ TR_RACK_DECAYGF=both TR_RACK_KNN=both TR_RACK_TMSELECT=both ./out/ModularBot The unit test `testRevertOverrideRestoresFullRack` exercises exactly this table. FLOOR PATH, verified not assumed: `chooseFromFit` already returns `admitted[0]` on the floor path, so it respects admission by construction. Cold field + shipped default -> floor returns Pattern (id 5), NOT HeadOn. With an explicit all-`both` membership the same cold field returns gun 0 (HeadOn) - the old behaviour. Four assertions in `testFloorRespectsAdmission`. LIVENESS: one 1-round battle with NO overrides -> Pattern selected 105/105 = 100%, every other gun 0 including TMPattern. Honesty caveat retained in the doc: 4 of the 10 opponents were Tank Royale sample-bot PORTS rather than the original classic jars (only DrussGT is a real classic jar through the shim). Guards: test_rack_membership 48 (was 38; new floor/revert/default checks), test_tm_pattern_registration 20 (5 checks hard-coded the old default and were updated to assert the new one, with the TMPATTERN parity proof moved onto an explicit old-rack table), test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 28, test_selector_tiebreak 19, test_tm_pattern_rack_live 4, test_tm_pattern_learning 3, acceptance_offline_vs_online 12/12 VERDICT PASS. ModularBot compiles. FOLLOW-ON THIS EXPOSED: membership filters SELECTION but not virtual-bullet SPAWNING, so under `onlyPattern` the 13 unselected guns still predict and spawn every tick. Tsetlin alone is ~5.3 ms/tick (~41% of the 13.16 ms per-tick budget), so we are still paying for it while never using it. Gating spawn on admission would reclaim that; it was deliberately NOT done here because it would alter the measurement protocol mid-A/B. |
||
|
SirStone
|
9cd6e9b8ce |
Ablation: the radial TM is replaceable by a CONSTANT, and its avenue is dead on the
shipped metric The radial TM beats Linear on bmPoint, but its head never beat the majority baseline after the label bias was fixed - suggesting the win is a constant lean rather than learning. So: sweep a stateless constant short-range offset (new `common_libs/guns/radial_offset.nim`, no learning at all) against the learned TM. VERDICT (measured, offline range, seeds=3, 18 paired runs, 231 TM rounds): 1. **bmPoint - REPLACE the TM with a constant.** `RO_s0.95` (aim distance x0.95) TIES it early (9/9, p=1.0) and BEATS it overall (15/3, p=0.0075; 7.47% vs 6.89% per-run mean). A fixed -20px does the same. The head never beats its majority baseline (56.2% vs 57.2%). 2. **bmPath (the SHIPPED metric) - the radial avenue is a DEAD END.** TMRadial is a systematic LOSS there (2/16, p=0.0013); every constant is within +-0.2pp; the only real bmPath effect is the BotRadius clamp. So the radial shift cannot help the shipped configuration. 3. The per-adversary optimum DOES vary (fixed -10 for crazy, -30 for tr_crazy, scale 0.95 for three others) - but ONE GLOBAL CONSTANT still beats the adaptively-trained head, so the "fragility justifies learning" argument FAILS. THE REAL FINDING UNDERNEATH, and it generalises beyond this gun: the base linear prediction systematically OVERSHOOTS. Measured raw per-tick base radial error has mean -71 to -100 px; the enemy is NEARER than the prediction in 63-81% of shots and farther in only 4-14%, CONSISTENT ACROSS ALL SIX CAPTURES. Radial label histogram [415166,126461,120325,44694,19351] = 57.2% majority class, mean label -82.3 px, mean applied shift -37.6 px. So the net-short bias is a GENUINE property of these range-holders against a constant-velocity extrapolation (they decelerate and turn, so the true position is closer than the straight-line guess) - NOT a fixture artefact. That is worth chasing for the guns that actually ship. Caveat: bmPoint is not the shipped metric (bmPath won the real-hit-rate A/B for SELECTION), so a bmPoint win is not yet evidence of a real win. That needs a live test - and the natural target is Pattern, which is now the default and best gun. Adds radial_offset.nim + sweep_radial_offset.nim; tm_pattern.nim gains additive instrumentation only (radial label mean and applied-shift mean; no behaviour change, and test_tm_pattern_registration still passes all 20 checks). |
||
|
SirStone
|
589a230106 |
TM radial gun: registered (default OFF) + label-bias fix that removes the bias but
retracts its own earlier learning claim === TASK 1: REGISTERED AS GUN 14, DEFAULT `off` === The radial TM gun is now a first-class rack member (`TMPATTERN`, id 14), forceable alone with `TR_RACK_TMPATTERN=both` plus every other `TR_RACK_*=off`. DEFAULT IS `off`, and the justification matters: `both` would let it compete for selection AND (because the shared VirtualTracker ring is order-sensitive) shift every other gun's learning order, so it CANNOT leave the default path unchanged. With `off` its predict and spawnBullets are additionally GATED on rack admission (the only gun wired that way), so the shipped default never spawns it at all: zero cost, zero ring perturbation. Live proof: 1-round battle with only TMPATTERN racked -> `gun 14 (TMPattern): vShots=400 selected=104 other-gun selections=0`. Default-path-unchanged proof: parity checks that the 15-gun default bestGun/ selectGun equals the old 14-gun rack RNG-draw-for-RNG-draw, that gun 14 is never selected by default, and acceptance 12/12. Cost: 0.36 ms/tick (predict 0.30 + onResult 0.05) ~= 3% of the 13.16 ms budget. Tsetlin in the same harness is 1.62 ms/tick, so the new gun is ~4.5x cheaper. === TASK 2: THE LABEL-BIAS FIX - AND A RETRACTION === Root cause confirmed: under bmPoint a SHORT radial correction resolves the virtual bullet BEFORE the base arrival tick, so the label was dropped (labelMisses). Fix: defer the label in a pending queue and flush it once the arrival tick is recorded; labels still come from the BASE arrival tick. labelMisses 4,281,695 -> 0 training samples 1,071,824 -> 5,345,847 (x5) radial head acc 48.8% -> 57.0% (shuffled control 20.0%) bmPoint hit rate 9.4/5.8% -> 9.1/5.7% (unchanged, within noise) So the fix IMPROVES LEARNING but NOT the metric. **RETRACTION OF THE PREVIOUS JOB'S CLAIM.** It reported the radial head's 48.8% against a 36.7% majority baseline and concluded "conditional learning, not a constant bias". With the bias removed, the correctly-measured majority baseline is **58.2%** - so the head at 57.0% is AT/BELOW majority. The earlier apparent conditional learning was PARTLY AN ARTEFACT OF THE BIASED SAMPLE. The bmPoint metric win is real (TMRadial > Linear early 16/2 p=0.0013, overall 18/0 p<0.0001; > shuffled 18/0 p<0.0001) but it comes from a NET-POSITIVE AVERAGE RADIAL SHIFT, not from beating a majority classifier. Recorded plainly rather than left standing. Guards: test_tm_pattern_registration 20 (new), test_tm_pattern_rack_live 4 (new), test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3 (the SIGSEGV is gone - the knn_gun rewrite is now committed), test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 28, test_rack_membership 38, test_selector_tiebreak 19, test_tm_pattern_learning 3, acceptance_offline_vs_online 12/12. ModularBot compiles (release). Note: `common_libs/tests/range_guns.nim` still builds 14 offline drivers (the offline sweep constructs TmPatternGun directly and acceptance only inspects ids 0..13), so nothing breaks - but a future job wanting it in the offline rack must add a 15th driver and mirror the live admission gating. gun_stats.jsonl now emits 15 rows; downstream tooling should ignore id 14. |
||
|
SirStone
|
394b3deeed |
SETTLED: Pattern alone is the best single gun IN GENERAL, not just vs DrussGT
Closes the one-adversary caveat that blocked shipping `onlyPattern`. Two arms
(`full` vs `onlyPattern`, rack knobs only), one frozen binary from clean HEAD
(
|
||
|
SirStone
|
a54ae6a162 |
SETTLED: no small rack beats Pattern alone; the selector is negative value on a GOOD rack
Follow-up to |
||
|
SirStone
|
4657fe715e |
wave pairing: 36-58% of GF/DecayGF/KNN learning samples were MISLABELLED
The audit inferred (from code) that GF/DecayGF/KNN pop the OLDEST wave on resolution, while under bmPath bullets leave the arena in NON-FIFO order - so an outcome could be attached to the wrong wave. It also noted that `starved=0` does NOT rule this out. Both halves are now MEASURED. MISPAIRING RATE (10 DrussGT fixtures, real VirtualTracker, 344k resolutions/gun): gun bmPath mispair label err bmPoint mispair label err GuessFactor 36.48% 19.39% 18.24% 7.62% DecayGF 36.85% 19.52% 20.57% 8.64% KNN 57.91% 27.63% 29.75% 11.58% (starved = 0 everywhere, exactly as the audit predicted) So ~1 in 5 GF/DecayGF learning samples and ~1 in 4 KNN samples carried a WRONG guess-factor bin. This is a material corruption of the learning signal. FIX: the same fireTick-keyed ring scheme `tsetlin.nim`/`tm_selector.nim` already use - `slot = (fireTick*4 + bin) mod 1024` (period 256 ticks, longer than the ~91-tick max flight), looked up by exact key. Public interfaces unchanged; added `waveResolved`/`waveMispaired` integrity counters. AFTER: mispaired = 0 and starved = 0, both metrics, all three guns. EFFECT ON HIT RATE: SMALL AND NOT SIGNIFICANT. bmPath 4000 samples/gun: GuessFactor 23.20% -> 23.02% (-0.18pp, per-run sign-flip p=0.750) DecayGF 23.80% -> 24.25% (+0.45pp, p=0.625) KNN 18.27% -> 18.80% (+0.53pp, p=0.547) bmPoint: +0.05 / +0.33 / -0.15pp, p = 1.00 / 0.50 / 0.50. Per-run ranges overlap almost completely. A bullet-level z-test is anti-conservative (bullets within a fixture share a trajectory) and its KNN p=1.9e-16 cannot be trusted given ~10 effective independent runs. PLAIN READING: this is a CORRECTNESS fix, not a measurable hit-rate win. It removes a 36-58% mislabelling of the learning signal; the point estimates move by at most ~0.5pp, within run-to-run noise. Stated plainly rather than oversold. A REGRESSION IT CAUGHT IN ITSELF (and this explains the SIGSEGV another job saw and correctly attributed to a concurrent knn_gun.nim rewrite): the first implementation put an inline `array[1024, KNNWave]` (~100KB) inside each gun, which overflowed the default 8MB stack and made `test_power_selection` SIGSEGV. Causation was proven by stashing only the three gun files (test passed), then fixed by making the rings heap-backed `seq`. Verified: `test_power_selection` 3 PASS on the default stack, and zero inline `array[1024]` remain. Guards: test_wave_pairing 17 (new, pure), test_gun_harness 39, test_vbullet_metric 11, test_power_selection 3, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 28. ModularBot compiles. Adds audit_wave_pairing.nim and compare_pairing.nim. |
||
|
SirStone
|
1ea72c7f14 |
TM gun round 2: base was never behind; RADIAL target beats Linear on bmPoint
=== TASK 1: MY PREMISE WAS REFUTED ===
I instructed the job to "fix the baseline" because an earlier measurement said the
TM gun's base did not iterate flight time like `LinearGun`. MEASURED: the new gun's
base is BYTE-FOR-BYTE `LinearGun` - 18/18 runs tie exactly, p=1.000, every per-run
row byte-identical. The "non-iterating baseline" belonged to the OLD `tsetlin.nim`,
not this gun. So no fix was needed, and the earlier inference should not have been
generalised to the new gun. (It did still align the zero-correction clamp to
LinearGun's exact [0, arena] range, and reports the old BotRadius-inset base was a
wash/marginally better at 34.2%/24.7%.)
=== TASK 2: THE RADIAL TARGET - A CONTROL-VALIDATED WIN, BUT ONLY ON bmPoint ===
Instead of the lateral (GF-bucket) component - which the linear lead already
captures - the TM now predicts the RADIAL component: will the enemy be nearer or
farther than the base prediction when our bullet arrives? A 5-class radial head
sharing the same 40-bit context and TM core; the readout advances/retards the aim
distance along the base bearing.
under bmPath (the SHIPPED metric): STRUCTURAL NO-OP
synthetic 8/8 exact ties, p=1.0; real 33.9%/24.1% vs Linear 34.0%/24.3%
under bmPoint: A WIN, control-validated
TMRadial 9.4% (6013/63785) / 5.8% (42079/726652)
Linear 7.2% / 4.7% overall 17/1, p=0.0001
Tsetlin 7.0% / 4.8% overall 15/3, p=0.0075
shuffled 7.0% / 3.6% early 17/1 p=0.0001; overall 18/0, p<0.0001
radial head online accuracy 48.8% vs 19.9% shuffled chance and 36.7% majority
-> it is CONDITIONAL learning, not a constant short-range bias.
Best config: TM_RADIAL_RANGE=60, TM_RAD_MARGIN=0.25, 5 classes.
CAVEAT THAT MATTERS: a win on `bmPoint` is NOT yet evidence of a real win. `bmPath`
is the shipped SELECTION metric precisely because it beat `bmPoint` on real hit
rate (7.43% vs 4.70%). But that A/B was about which gun to PICK, not about gun
QUALITY - a gun can be better in reality while scoring worse on the selection
metric. So this needs a LIVE test, and it is the decisive one.
=== TASK 3: REVERSAL TARGET - CLEAN NEGATIVE ===
The label positive rate is only 9.7% (rev=[24772,2673]) and the head's 86.8%
accuracy is BELOW the 90.3% majority baseline: it does not learn the positive
class at all. Hit-rate effect neutral (bmPath 19.5%/18.4% vs shuffled 19.1%/17.8%,
p=0.24/0.82). Dropped.
=== OVERALL ===
Not competitive on the shipped bmPath metric (gated GF 28.3%/22.2% vs Linear
34.0%/24.3%, p=0.0075). Better than Linear on bmPoint via TMRadial (+2.2pp early,
+1.1pp overall). Per-enemy reset exists; a fresh gun per round; NO cross-battle
persistence (the user's non-negotiable).
MEASURED LIMITATION: radial mode has a high labelMiss because aiming short
resolves BEFORE the base arrival tick, biasing training toward resolvable samples.
The metric win is label-independent. A deferred-label fix is the next refinement.
INFERRED: the mechanism is surfers being NEARER than the base prediction
(range-holding); a constant-short-offset ablation would separate a learned
short-range bias from genuine per-tick conditional prediction.
|
||
|
SirStone
|
78975a35c4 |
cost: parallel per-enemy virtual bullets are NOT affordable as proposed
Benchmark driving the real rack and the real VirtualTracker over 7 recorded DrussGT fixtures synthesised into an N-enemy melee. Answers "the virtual bullets are cheap, why not keep fitness for every enemy in parallel?" (the user's idea, motivated by making kill-stealing target switches free). BASELINE: exactly 4.0 predict calls per gun per tick (one per power bin) - 52/tick for the shipped 13-gun rack (TMSelect is compiled out). The task's 56/tick was the 14-gun figure. VERDICT: NOT AFFORDABLE. Budget is 13.16 ms/tick (76 ticks/s measured live). N=1 46% of budget N=2 94% <- already at the edge N=4 189% N=6 274% Marginal cost ~= 5.9 ms per extra target, linear. TWO FINDINGS THE PROPOSAL MISSED: 1. `onResult` TRAINING dominates, not predict. Tsetlin's onResult alone is 3.40 ms/tick - ~99.5% of all 13-gun onResult cost - doing ~174k rand() calls per resolved bullet. Every spawned bullet that resolves triggers it, so it scales 1:1 with targets. The per-target cost is the Tsetlin training pass. 2. `MaxBullets=8192` is a HARD BLOCKER, not just CPU. Spawn rate is 56*N/tick and path-metric bullets live until they hit a wall (40-90 ticks). Measured dropped bullets/tick: N=1 -> 0, N=2 -> ~3, N=4 -> ~180, N=6 -> ~300. At N=6 the ring wraps every ~24 ticks, so most bullets are silently clobbered and never scored. A working N=6 pipeline needs MaxBullets ~30k-50k (~4-6 MB, cheap RAM). ALSO MEASURED: Tsetlin and KNN do NOT cache per tick - they redo the full TM forward pass / full KNN scan for EACH of the 4 power bins (Tsetlin 1.94 ms/tick of predict, KNN 0.45). The earlier "tick-only cache" fix never touched the two most expensive predicts. Pattern and TMSelect do cache fully. MITIGATIONS (measured predict+spawn at N=6 vs 13.59 ms baseline): nearest-K=1 only 45% budget nearest-K=2 95% rotate every 3 ticks 95% drop Tsetlin for extras ~68% (INFERRED from Tsetlin's measured 90% share) Tsetlin is ~90% of the per-target cost, so excluding it from non-primary targets makes N=6 fit. "Resolve less often" is not a separate lever - resolution IS when training happens. ARCHITECTURAL CAVEAT (correctness, not cost - and not priced into the proposal): the shared-rack topology is broken for this. The guns are global singletons, so predicting for enemy B ADVANCES/OVERWRITES enemy A's velocity tracker, KNN feature history and Tsetlin frame window in the SAME instance. Per-enemy fitness with correct histories therefore requires PER-ENEMY GUN INSTANCES, which is what this benchmark measured. That multiplies the (already dominant) Tsetlin cost. CONSEQUENCE FOR THE PLAN: combined with the measured finding that the selector is negative value and the rack should shrink to a few good guns, this work is much less valuable than assumed - with a small rack (Pattern's predict is 40us and fully cached) the cost falls proportionally. Priority lowered accordingly. Caveat: the host was heavily loaded (load 15/16), so absolute ms carry ~30-50% noise; min-of-2 and two independent runs agree on the trend, the Tsetlin dominance, and the ring overflow. No melee fixture exists in the repo, so the 7 enemies are 7 distinct recorded trajectories (stated in the file header). |
||
|
SirStone
|
e0666a562d |
The gun selector is NEGATIVE value: Pattern alone beats the full rack (p=0.0012)
MEASURED against the real DrussGT, one frozen binary built from clean HEAD, rack knobs only (no source edits), 5 arms x 7 runs x 7 rounds, 8 concurrent battles, judged ONLY on server-side real hit rate from the events sidecar, exact two-sided permutation test on per-run rates. arm runs shots hits real % dmg/run p vs full full (shipped) 7 3898 270 6.93 159 -- onlyPattern 7 4582 494 10.78 287 0.0012 <- BETTER onlyKNN 7 4033 207 5.13 119 0.1340 onlyLinear 7 3215 105 3.27 65 0.0082 onlyGF 7 3193 72 2.25 45 0.0012 Firing Pattern ALONE gives +3.85pp pooled hit rate and +80% damage per run, and it fires MORE shots (4582 vs 3898) - it dominates on rate and volume. This is not "any single gun wins" (full beats Linear, GF and KNN); it is specifically "Pattern alone beats the rack". WHY - the virtual fitness signal mis-ranks guns against real outcomes: - HeadOn is massively over-selected: 31.4% of ticks, the most real shots (1070), but only 4.5% REAL. It alone drags the rack down. - Pattern has the best virtual rank and near-best real rate (11.9%, rank 2), yet is selected only 22.6% of the time. - Linear's apparent strength was SELECTION BIAS: conditional on being selected it looked like 15.2% (n=33), but its UNCONDITIONAL rate (onlyLinear) is 3.27%. Every earlier per-gun "real rate" in this repo is conditional on selection and is therefore confounded. This experiment is the clean measurement. NOT YET SETTLED (do not overclaim): - ONE ADVERSARY. All of this is vs DrussGT. Pattern must be re-checked against other bots before it becomes the default on this evidence alone. - Whether a SMALL rack of good guns beats Pattern alone. The selector is negative value on the CURRENT bloated rack; that does not prove it is negative value on a rack of only good guns. That is the next experiment and it decides whether the selection apparatus is fixed or disabled. - The user's standing directive is to KEEP virtual-fitness selection. This measurement conflicts with it, so the next step tests the selector on a small good rack rather than assuming either answer. Context - three prior selection-side attempts all failed: hysteresis (7.02% -> 5.10%, p=0.002), commitment (7.17% -> 4.44%, p=0.0012), arrival-accuracy tie-break (7.08%, p=0.88 null). The per-tick random draw is load-bearing on three independent measurements. This experiment locates the real problem one level up: which guns are in the rack, and that the virtual signal ranks them wrongly. Preserves the reusable harness (tools/ab/which_gun_run_one.sh, which_gun_arm_env.sh, which_gun_analyze.py) and the full writeup (docs/selector_negative_value.md). |
||
|
SirStone
|
0ede6d12ec |
selector: arrival-accuracy tie-break measured NEGATIVE; randomness is load-bearing
Hypothesis under test (from the gun audit, which named the tie-band as "the lever that matters most"): `bmPath` is deliberately generous (2.3-3.6x `bmPoint`), so a gun can sit in the tied band on a ray that sweeps the target's path while its bullets ARRIVE badly. So: keep the `path`-ranked band (path beat point on real hit rate 7.43% vs 4.70%, z=5.56), but narrow the random draw inside it using a parallel `point` (arrival-accuracy) window. RESULT: NO EFFECT. Real DrussGT, ONE frozen binary (/tmp/ModularBot_tieband, md5 2c0c56e6...), env knobs only, 7 runs x 7 rounds per arm, server-side events sidecar, exact two-sided permutation test on per-run rates. arm runs shots real % dmg/run d p tbbase (shipped) 7 4128 7.17 175 -- -- tbpt path-rank + point-narrow 7 3938 7.08 165 +0.14 0.88 tbpc =commit control 7 3759 4.44 98 +2.74 0.0012 tbpt25 point margin 0.25 7 3683 5.59 119 +1.65 0.20 tbtie05 / tbtie40 (band width) 7 3937/3917 5.84/6.28 133/144 1.49/1.00 0.11/0.25 tbwin50 (SelectorWindow=50) 7 3983 6.05 139 +1.20 0.11 tbfloor10 (FloorPeakFrac=0.10) 7 3829 5.33 118 +2.12 0.11 tbpt vs base: fully overlapping ranges, p=0.88. This is a REAL null, not a dead arm - the mechanism was live, and it visibly changed the selected-gun mix (Pattern 24%->16%, Accel 6%->16%, Tsetlin ~0%->13%). CONTROL VALIDATED, AND THIS IS THE THIRD TIME: removing the random draw inside the band is SIGNIFICANTLY WORSE (4.44%, p=0.0012). Combined with the earlier hysteresis A/B (7.02% -> 5.10% for commitment) and the light-hysteresis result, the selector's per-tick randomness is now load-bearing on three independent measurements. Narrowing the band on ANY second virtual statistic has not helped. Every knob swept (band width, floor, window) is nominally worse than shipped at n=7; that is "no credible win" rather than "proven harm" (sd ~1.8pp, ~1pp resolution, underpowered). Shipped default stays `GUN_SELECTOR_TIEBREAK=off`; the feature is opt-in, fully guarded, and costs zero extra work on the default path (point windows are scored only when the mode is on). Guards: test_selector_tiebreak 19 (new, pure), test_gun_harness 39, test_vbullet_metric 11, test_adaptive_radar 41, test_tfil_ring_weights 24, test_power_policy 26, test_ram_decision 28, test_rack_membership 38, acceptance_offline_vs_online 12/12 PASS (offline path calls neither chooseFromFit nor the tie-break). STRATEGIC CONCLUSION: three selection-side attempts have now failed (hysteresis, commitment, point tie-break). The selector is at a local optimum and the remaining lever is the QUALITY OF THE GUNS, not the selection among them. |
||
|
SirStone
|
ca82053a11 |
TM gun: the discrete-target diagnosis was RIGHT - it learns now. Still loses to Linear.
The user's goal: a TM gun that is the best 1v1 gun, starting from scratch every
battle but quickly overfitting the current enemy. The previous attempt (knob
tuning) failed: NO configuration beat its own shuffled-feedback control, and the
TM-off ablation scored the same as TM-on, i.e. the TM's correction was
near-zero-mean noise. Diagnosis then: a Tsetlin Machine is a CLASSIFIER, and we
were asking it for an absolute aim point - a regression target. So this attempt
gave it a DISCRETE target (multi-class over guess-factor buckets) with 40
binary/bucketed motion features, and measured it against Linear, the default
Tsetlin gun, and a MANDATORY shuffled control.
THE DIAGNOSIS IS CONFIRMED - THE TM LEARNS, DECISIVELY:
online class accuracy 46.0% vs shuffled control 20.0% (2.3x chance)
raw ungated argmax 21.2%/18.6% vs shuffled 15.2%/8.3% (18/18, p<0.0001)
TMPattern > its shuffled control, overall 17/1 runs, p=0.0001
Compare the previous attempt, which could not beat shuffled feedback at all.
TMPattern also beats the default Tsetlin gun early (17/1, p=0.0001), so it is a
strictly better TM gun than the one in the rack.
BUT IT IS NOT COMPETITIVE WITH LINEAR ON REAL SURFERS:
real DrussGT, bmPath (the shipped metric), 3 seeds, pooled early/overall
Linear 34.0% (6358/18715) 24.3% (58297/239943)
TMPattern (gated) 27.9% (15514/55535) 22.0% (158658/719681)
TMPatternShuf 28.7% 19.4%
Linear > TMPattern: 15/18 early p=0.0075, 15/18 overall p=0.0075
bmPoint: neutral (7.2%/4.6% vs Linear 7.2%/4.7%)
synthetic controlled motion: matches/edges Linear (66.8%/60.6% vs 66.4%/59.6%,
shuffled 55.7%/50.1%) - the mechanism works when motion is predictable.
So: the representation fix moved this from "learns nothing" to "learns strongly
but applies its knowledge badly". INFERRED reason for the residual loss: the
linear lead is already the modal GF bucket (the label histogram is centred), so
corrective excursions away from it are net-negative. The measured deficit lives
in the BASELINE and in RANGE, not in the TM knobs - which is why further knob
tuning was never going to work.
Best config: gated hard K=5, TM_CONF_MARGIN=0.25, TM_SHRINK=0.5.
NOT TRIED (time-boxed): the binary-reversal target, and a RADIAL (range-holding)
target - the latter is the top next step.
Adds `common_libs/guns/tm_pattern.nim` (NOT registered in the rack),
`common_libs/tests/sweep_tm_pattern.nim`, and a durable writeup at
`common_libs/tests/tm_pattern_sweep_results.md`.
|