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`.
The user's goal was "a TM gun that can learn fast and generalize better".
Swept offline over the real DrussGT fixtures (no live battles) by coordinate
descent, one lever at a time, with a SHUFFLED-FEEDBACK CONTROL - a TM trained
on randomised targets. That control is what settles the question.
Final confirmation, 4 seeds each (~74,600 first-100-tick bullets per config):
config EARLY(first 100) OVERALL
Shuf_w3 (RANDOM feedback) 23.9% 20.0%
win3_s1.1 (best real TM found) 23.7% 20.2%
Shuf_w10 (RANDOM feedback) 23.1% 20.0%
win3_st100 (prior job's edit) 23.0% 20.1%
win3_off (TM correction ~= 0) 22.7% 20.2%
def_w10 (shipped default) 22.1% 20.3%
Linear (deterministic reference) 34.0% 24.3%
The best real config beats the default early (23.7% vs 22.1%, non-overlapping
per-seed ranges, z=+7.34, p=2e-13) - but its own SHUFFLED control scores 23.9%,
i.e. HIGHER, z=-0.91, p=0.37. Random targets do at least as well. So the early
gain is not learning.
Per-lever screens were flat: TM_N_CLAUSES 25/50/100/200 all 23.0% early,
completely flat; TM_N_STATES 4/32/100 all ~22-23% (unstable across seeds);
TM_S mildly monotonic (lower better early); TM_T flat; TM_WINDOW_SIZE 2/3/10
all within noise of each other and of the shuffled control.
Two further findings:
- The TM-off ablation (correction ~= 0) scores 22.7%/20.2%, essentially the
same as TM-on. The TM's correction is near-zero-mean noise; the gun's
one-shot internal linear baseline accounts for its accuracy.
- The TM gun is 10.3 pp behind Linear early and 4.1 pp behind overall. That
deficit is in the BASELINE MODEL (LinearGun iterates flight time; this gun
does not), not in the TM hyper-parameters. Tuning knobs cannot close it.
Conclusion: do not tune TM hyper-parameters further. Either the input
representation or the prediction target is what needs to change - the shuffled
control shows the TM is not extracting target information beyond its baseline.
Defaults left UNCHANGED (window=10/states=32/S=1.5/T=25/clauses=50); an
uncommitted prior edit (window=3/states=100) was reverted as unsupported.
Hyper-parameters are now compile-time overridable (-d:TM_WINDOW_SIZE=3 etc.)
so future sweeps need no gun edit.
NOT MEASURED: real hit rate vs DrussGT (offline only by design). The repo's own
docs/gun_rack_analysis.md 2 reports the virtual metric is a sign-unstable ranker
of real hit rate, so the comparison against "Linear 10.7% real" is not direct -
whether the TM is competitive live is INFERRED-unknown, not measured.
Guards: test_gun_harness 39/39, test_vbullet_metric, test_power_selection,
test_tsetlin_gun, test_tm_pattern_learning all green.
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.
The previous fix (learn the residual against a constant-velocity base) was
structurally right but cost us on real wave-surfing movement: GF 108 -> 55,
KNN 101 -> 74 on the classic DrussGT captures. Root cause: the linear base is
a poor model for a surfer, so the residual histogram is noisier than the old
total-lead histogram.
FIX: blend the RANGE between a radial-only forecast and the geometric one by
radialFrac (the fraction of recent per-tick motion that is radial), keeping
the constant-velocity bearing. dist = radialDist + rf*(linearDist - radialDist).
New VelocityTracker in common_libs/guns/lead_forecast.nim; the window default
is 32 and results were identical at 16 and 40, so it is not tightly tuned.
Nine candidate bases were measured and rejected WITH NUMBERS rather than by
argument, which is why I trust the winner:
velocity scaling 0.8 recovers DrussGT but destroys wall-bounce 241 -> 20
radial-only range excellent DrussGT, wall-bounce 241 -> 140
short-window averaged vel worse than both bases outright
hard reversal/speed gates help DrussGT, lose nothing, but weaker than blend
radial-fraction blend best on BOTH <- shipped
Result (hits per 2000; classic-5 = classic DrussGT captures, tr-5 = the new
closed-loop TR captures, synth-10 = the rest):
base classic-5 GF/DGF tr-5 GF/DGF synth-10 GF/DGF
current(prefix) 108 / 108 41 / 39 1302 / 1302
linear(postfix) 55 / 76 9 / 4 2702 / 2692
BLEND 171 / 100 86 / 87 2717 / 2703
Strictly better than both on classic-5 GF and on every synthetic bucket. The
one figure below the old base is classic-5 DecayGF (108 -> 100, -8/2000,
within noise) and that is stated plainly rather than hidden.
TASK B - enemy energy in learners. KNN gains an 8th feature, enemyEnergy/100,
on a FIXED [0,1] scale (not min-max) because threshold behaviour keys off
absolute energy. Honest result: it is NEUTRAL on the target fixture (77 vs 77)
and roughly neutral in aggregate. The base change, not the feature, moved that
fixture. Tsetlin already encoded enemyEnergy and now scores 88/400 on
energy-threshold-turner against Linear's 43/400 - a 2x margin, which is the
'can a TM learn a high-level pattern' question answered in gun form.
TASK C - is the virtual-bullet metric itself faithful? Quantified: scoring the
bullet's PATH against BotRadius instead of the single point at aim distance
raises every gun by +31% (GF) to +86% (HeadOn), so the current model is
PESSIMISTIC, and it RE-RANKS materially: Linear 9th -> 6th, AvgLead 7th -> 3rd,
GuessFactor 4th -> 9th, DecayGF 6th -> 12th. The 12/12 offline==online
acceptance still holds under the path model (verified with a temporary env
hook driving both sides), so no red flag. VERDICT: do NOT switch. The point
model is the standard virtual-bullet PREDICTION-ACCURACY fitness - the bullet
must arrive at the predicted point at the right time - while the path model
measures hypothetical hit chance against a target that never dodges, and in
open-loop fixtures it over-credits directional guns (HeadOn 35% on DrussGT,
100% on constant-velocity) for exactly that reason. The models differ
materially but the current one is not shown to be unfaithful FOR ITS PURPOSE.
Because the metric drives gun SELECTION, this is now being A/B'd against real
hit rate versus the live DrussGT boss, which is the only ground truth we have.
Verified: 20 fixtures 35636/104000 (34.3%); 33 guard checks; 12/12 acceptance;
tsetlin tests green; live gauntlet 5/5.
The entire GuessFactor family scored 0% on clean circular and wall-bounce
trajectories. Two hypotheses were on the table and BOTH were wrong:
- MEA range too narrow / edge clamping: REFUTED. Measured 0 clamped shots
out of 837/849/957, required offsets peak at ~33 deg against MEA
28.1-46.7 deg, and the 8 in arcsin(8/bulletSpeed) is correct (it is the max
robot SPEED, not the hit radius). Changing it to BotRadius=18 would have
coarsened resolution for nothing.
- Peak selection: REFUTED. A sweep of every constant GF value showed the
ORACLE-BEST constant offset on the original gun was only 6% circular,
4% wall-bounce, 7.5% random-walk. No peak choice could have done better.
The learning path was fine too: ~850-960 observations per fixture, 0
starved waves, well-populated histograms.
REAL CAUSE: the GF family aimed at the FIRE-TIME distance. The virtual-bullet
metric resolves a bullet at the AIM-POINT distance and scores that single
point against the enemy's position on that tick, so with any radial target
motion the bullet stops at the wrong radius and misses even with a perfect
angle. Angle-only prediction is structurally unscoreable under this metric.
FIX: give the GF family a self-consistent constant-velocity forecast as its
base reference (new common_libs/guns/lead_forecast.nim, which iterates the
flight time to the same fixed point circular.nim uses), so the histogram
learns the RESIDUAL against that forecast and the aim point lands at the
right radius. Applied to guess_factor, decay_gf and knn_gun.
Same defect fixed in Linear: it did a one-shot dist/bulletSpeed extrapolation
and never iterated its flight time.
The oracle sweep proves the structural fix, independently of tuning: the best
achievable constant GF moved 6% -> 20% (circular), 4% -> 57% (wall-bounce),
7.5% -> 49% (random-walk).
MEASURED, all 15 fixtures: total 39.0% -> 44.4% (30399 -> 34654 hits).
circular GF 6 -> 23, DecayGF 6 -> 21
wall-bounce GF 0 -> 60.2, DecayGF 0 -> 60.2
constant-vel GF 26 -> 100, DecayGF 26 -> 100, KNN 26 -> 100, Linear 87 -> 100
random-walk GF 0 -> 53, DecayGF 0 -> 52, Linear 24 -> 53
StraightLine GF 8 -> 77, DecayGF 8 -> 77
Non-regression: 33 guard checks pass, the range's 12/12 offline==online
acceptance still PASSES, tsetlin tests green, live gauntlet 5/5.
HONEST TRADE-OFF, recorded rather than hidden: on the 5 real DrussGT
wave-surfing captures the GF family REGRESSES - GuessFactor 108 -> 55,
DecayGF 108 -> 76, KNN 101 -> 74 hits per 2000. The linear base is a poor
model for a surfer, so the residual histogram is noisier than the old
total-lead histogram. Linear itself improved there (95 -> 105). The synthetic
range and the live gauntlet both improved, and the structural bug is provably
fixed, so this was judged worth the cost - but recovering the DrussGT
regression is the next job, not something to wave away.
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.
Four guns cached a whole prediction per tick while predict() is called once
per power bin, so every bin after the first (and the real fired shot, which
shares lastState) reused the power-1.0 lead. Fixed by caching only the
speed-INDEPENDENT derived state and recomputing the lead per requested speed:
- stop_shot: also fixes prevSpeed being written before it was read, which
made abs(speed) < abs(prev) permanently false and the entire
stop-prediction branch unreachable (it was just Linear).
- displacement: the cache key included bulletSpeed, so the guard missed on
all four bins and the 15-tick window advanced ~4x/tick, making the
inferred velocity ~4x too small.
- averaged_lead: tick cache removed outright. pattern_matcher: split into
speed-independent match+path and per-call lead.
FeedbackEvent gains fireTick/powerBin (additive; only virtual_bullets
constructs one) so guns can pair feedback to the exact shot instead of
guessing by coordinates. tsetlin uses it: traces are now keyed exactly by
(fireTick, powerBin) with a 1024-slot ring, and the 10-frame window shifts
at most once per tick (it was shifting ~4-5x/tick, so isWarmedUp tripped
after ~2 ticks).
KNOWN INCOMPLETE: tsetlin still does not diverge from Linear in battle. The
two named bugs are fixed (a 600-tick sim shows trainedShots=2141,
traceMisses=0, and a fixed-input probe converges to a 9.6px correction), but
the TM's clause feedback itself is broken: ~131 of 1740 literals end up
included per clause, so its conjunction never fires. Sweeping TM_S,
TM_N_CLAUSES and a two-branch Type-I update did not change the correction
from 0. Needs a real TM fix or removal, not another bug fix.
First-ever guard tests for the gun selector: common_libs/tests/
test_gun_harness.nim (14 checks, headless, no Java). There were none before,
which is how six broken guns survived a full analysis cycle. Against the
previous HEAD, 5 of these checks FAIL - that is the regression guard.
Wave queues (guess_factor, decay_gf, knn_gun): predict() stored ONE wave
per tick while onResult() popped one per resolved bullet (~4/tick), so the
queue drained to empty within a few dozen ticks, ~3 of every 4 resolutions
returned without learning, and the survivor paired with a same-tick wave
(bearingDelta ~= 0) pinning the histogram at centre. PROOF: GF.vHits ==
HeadOn.vHits and DecayGF.vHits == HeadOn.vHits byte-for-byte in every one
of 50 rounds — the guns had degenerated to HeadOn.
Now each gun keeps a per-bin FIFO with an O(1) head cursor. At most one
push per (tick, bin) so the fire site's 5th predict() call is a no-op, and
onResult pops the oldest wave of its OWN bin via e.bulletPower. Aiming
math untouched (it was already correct: 0 deg = East, CCW+).
maxBullets 2048 -> 8192: the rack spawns 52 bullets/tick so the ring wrapped
every ~39 ticks while a long power-3 shot needs ~90, silently discarding
unresolved bullets and biasing every measured hit rate by range. Added a
droppedBullets counter so a future overflow is measurable, and wavePushes/
waveStarved counters on the three guns. After the fix: vDropped = 0 and
vStarved = 0 across all 48 recorded rounds.
fitnessFor is now exported, deterministic (enemies iterated in ascending id
order) and shared by the selector and the stats dump, replacing a hand-rolled
merge in ModularBot that never advanced its window head.
Round lines gain additive keys: vDropped, vStarved.
- 30-tick cooldown after ghost-stuck/timeout ram exit prevents re-entry loop
- enemy_tracker.update() skips dead bots to prevent same-tick scan resurrection
- TFIL graphics cleared when ramming is active movement
- [config] logs: white base with green-highlighted changes only
- [ram:enter] logs trigger reason and key values on false→true transition
- [death] and [target-invalid] logs retained for diagnostics
New DrussGT-inspired modules:
- KNNGun: K-nearest-neighbor statistical targeting using GF density peaks
- GunheatTracker: dual-heat system (predicted + confirmed) for 1-2 tick lead
- ShadowTracker: computes GF regions safe from in-flight bullets (enemy wave dodge)
VirtualBodyTracker now integrates gunheat for earlier fire detection and shadows
for safe-zone multiplier (90% reduction in danger zones).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Cold-start bug: uniform bins[0..30]=0.1 made peakBin() always return
0 (first-wins tie), giving GF=-1 (max CW escape) before any learning.
Fixed with a triangular head-on bump at bin 15 (GF=0) as the prior.
- onResult now recomputes mea from FeedbackEvent.bulletPower instead of
the stale first-bin mea cached by predict; correct per-power-bin GF.
- Add DebugGF const (default false) with [gf-dbg] echoes in predict/onResult.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs fixed:
1. Multi-tick gaps: turn rate assumed 1 tick between observations, but scans can be 5+ ticks apart. Now divides by actual tickDelta.
2. Per-power-bin state corruption: predict() called 4x per tick (per power bin). After first call, prevHeading was already updated, causing subsequent calls to compute 0° delta. Now captures oldHeading/oldTick before updating.
Verified: OscillatorBot at 4°/tick captured correctly; normalization [-180°,180°] works; tickDelta=1 typical; first bin gets delta, subsequent bins see 0 (expected).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
SirStone