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.
This commit is contained in:
2026-09-22 01:33:31 +02:00
parent 1ea72c7f14
commit 4657fe715e
6 changed files with 586 additions and 65 deletions
+56 -27
View File
@@ -1,6 +1,16 @@
## Guess-factor gun: statistical targeting via GF histogram.
## Bins: 31, ranging GF -1 (max CW escape) to +1 (max CCW escape).
## Learns from virtual bullet outcomes; queues one wave per (tick, power bin).
##
## Wave pairing is EXACT: a wave is stored in a ring slot keyed by
## (fireTick, powerBin) and `onResult` looks up the wave with the resolution
## event's `fireTick`, NOT the oldest queued wave. Under the shipped bmPath
## metric bullets leave the arena in non-FIFO order (the aim direction changes
## every tick), so FIFO pairing attached outcomes to the wrong wave. Measured
## over the committed DrussGT fixtures, FIFO mispaired 36.5% of resolutions
## (19.4% of which changed the recorded GF bin). See
## common_libs/tests/audit_wave_pairing.nim. The exact key is the same pattern
## guns/tsetlin.nim and guns/tm_selector.nim use.
import std/[math, strformat]
import gun_harness/gun_interface
@@ -11,31 +21,43 @@ const
GFBins = 31
GFPrior = 0.1
DebugGF* = false
WaveCompactAt = 64 ## compact a bin's wave seq once this many entries are consumed
WaveRingSlots = 1024
## (fireTick, powerBin) -> ring slot. Period = WaveRingSlots/bins = 256 ticks.
## A bmPath bullet leaves an 800x600 (Tank Royale max 1000x1000) arena within
## ~91 (128) ticks, so a live wave is never overwritten by a newer one.
## Identical sizing to tsetlin.nim's TM_TRACE_SLOTS.
type
Wave = object
fireX, fireY: float
fireBearing: float # atan2(enemyY-selfY, enemyX-selfX) at fire tick (rad)
fireTick: int # key part: tick the bullet was fired
bin: int # key part: power bin the bullet belonged to
alive: bool
# mea not stored — recomputed from FeedbackEvent.bulletPower at resolution time
GFGun* = object
bins: array[GFBins, float]
# One wave queue per power bin. The owning bin is fixed at push time (from the
# bulletSpeed argument) and at pop time (from FeedbackEvent.bulletPower), so a
# resolved bullet is always paired with a wave from its own bin.
waves: array[len(vb.PowerBins), seq[Wave]]
waveHead: array[len(vb.PowerBins), int] # O(1) pop cursor into waves[bin]
# Exact (fireTick, powerBin)-keyed ring. A resolved bullet is matched to the
# wave it actually fired, no matter how many other shots resolved first.
# Heap-backed (seq) so the gun value stays small on the stack — a 1024-slot
# inline array overflowed the default 8 MB stack in test_power_selection.
waves: seq[Wave]
waveStoredTick: array[len(vb.PowerBins), int] # last tick a wave was queued for this bin
vt: VelocityTracker # enemy velocity history (base selection)
cachedTick: int # last tick the velocity tracker was advanced
wavePushes*: int # total waves enqueued (== one per (tick, bin))
waveStarved*: int # onResult found an empty queue for its own bin
waveStarved*: int # onResult found no live wave for its (fireTick, bin)
# ── pairing integrity ───────────────────────────────────────────────────
waveResolved*: int # onResult calls that found their exact wave
waveMispaired*: int # ring-slot collision (impossible by design): the
# slot held a different fireTick
debugGraphics*: bool
proc initGFGun*(): GFGun =
result.debugGraphics = false
result.cachedTick = -1
result.waves = newSeq[Wave](WaveRingSlots)
for b in 0..<len(vb.PowerBins):
result.waveStoredTick[b] = -1
# Seed with a head-on prior: triangular bump at bin 15 (GF=0).
@@ -51,7 +73,7 @@ proc gfToIndex(gf: float): int {.inline.} =
proc indexToGF(idx: int): float {.inline.} =
float(idx) / float(GFBins - 1) * 2.0 - 1.0
proc peakBin(g: GFGun): int =
proc peakBin*(g: GFGun): int =
var best = 0
for i in 1..<GFBins:
if g.bins[i] > g.bins[best]:
@@ -73,18 +95,9 @@ proc binForPower(power: float): int {.inline.} =
return i
-1
proc takeOldestWave(g: var GFGun, binIdx: int): (bool, Wave) =
## Pop the oldest unresolved wave for this bin (O(1) amortized via waveHead).
## Returns (false, default) when the bin's queue is empty.
if binIdx < 0 or g.waveHead[binIdx] >= g.waves[binIdx].len:
return (false, Wave())
result = (true, g.waves[binIdx][g.waveHead[binIdx]])
inc g.waveHead[binIdx]
# Amortized O(1): drop the consumed prefix once it dominates the queue.
if g.waveHead[binIdx] >= WaveCompactAt and
g.waveHead[binIdx] * 2 >= g.waves[binIdx].len:
g.waves[binIdx] = g.waves[binIdx][g.waveHead[binIdx] .. g.waves[binIdx].high]
g.waveHead[binIdx] = 0
proc waveSlot(fireTick, binIdx: int): int {.inline.} =
## Exact (fireTick, powerBin) key -> ring slot (same scheme as tsetlin.nim).
((fireTick * len(vb.PowerBins)) + binIdx) mod WaveRingSlots
proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPrediction =
if bulletSpeed <= 0.0:
@@ -104,10 +117,14 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
# call for the selected bin lands on the same tick and reuses the queued wave.
let binIdx = binForSpeed(bulletSpeed)
if binIdx >= 0 and g.waveStoredTick[binIdx] != state.tick:
g.waves[binIdx].add Wave(
let slot = waveSlot(state.tick, binIdx)
g.waves[slot] = Wave(
fireX: state.selfX,
fireY: state.selfY,
fireBearing: f.bearing,
fireTick: state.tick,
bin: binIdx,
alive: true,
)
g.waveStoredTick[binIdx] = state.tick
inc g.wavePushes
@@ -119,7 +136,7 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
let py = state.selfY + sin(gfAngle) * f.dist
when DebugGF:
echo fmt"[gf-dbg] predict: peakGF={peakGF:.2f} peakBin={peak} mea={radToDeg(mea):.1f}° aimAngle={radToDeg(gfAngle):.1f}° waves={g.waves[binIdx].len}"
echo fmt"[gf-dbg] predict: peakGF={peakGF:.2f} peakBin={peak} mea={radToDeg(mea):.1f}° aimAngle={radToDeg(gfAngle):.1f}° tick={state.tick}"
GunPrediction(
x: clamp(px, BotRadius, state.arenaWidth - BotRadius),
@@ -127,15 +144,25 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
)
proc onResult*(g: var GFGun, e: FeedbackEvent) =
## Called when a virtual bullet resolves. Pop the OLDEST unresolved wave from
## this bullet's own power-bin queue, compute the actual GF, and smooth-add it.
let binIdx = binForPower(e.bulletPower)
## Called when a virtual bullet resolves. Look up the wave with this event's
## exact (fireTick, powerBin), compute the actual GF, and smooth-add it. A
## missing key is counted, never silently mislabelled.
let binIdx =
if e.powerBin >= 0 and e.powerBin < len(vb.PowerBins): e.powerBin
else: binForPower(e.bulletPower)
if binIdx < 0: return
let (found, w) = g.takeOldestWave(binIdx)
if not found:
let slot = waveSlot(e.fireTick, binIdx)
var w = addr g.waves[slot]
if not w.alive:
inc g.waveStarved
return
if w.fireTick != e.fireTick:
# Ring slot collision: impossible while the ring period exceeds max flight.
inc g.waveMispaired
inc g.waveStarved
return
inc g.waveResolved
# Recompute mea from the actual bullet power (correct per-bin, not the cached first-bin mea)
let speed = bulletSpeed(e.bulletPower)
@@ -160,3 +187,5 @@ proc onResult*(g: var GFGun, e: FeedbackEvent) =
for i in 0..<GFBins:
let dist = abs(i - centerIdx)
g.bins[i] += 1.0 / float(1 + dist)
w.alive = false