fix(guns): per-bin wave queues unbreak GF/DecayGF/KNN learning; fix vbullet drops

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.
This commit is contained in:
2026-09-20 22:27:52 +02:00
parent a90a0cc9b5
commit 0cc682152d
5 changed files with 205 additions and 75 deletions
+16 -19
View File
@@ -199,24 +199,10 @@ method onRoundEnded*(bot: ModularBot, e: RoundEndedEventForBot) =
## Dump per-gun virtual bullet stats to /tmp/gun_stats.jsonl (one line per round). ## Dump per-gun virtual bullet stats to /tmp/gun_stats.jsonl (one line per round).
# Use lastKnownTargetId: currentTargetId is -1 if enemy died before round end # Use lastKnownTargetId: currentTargetId is -1 if enemy died before round end
let targetId = if bot.currentTargetId >= 0: bot.currentTargetId else: bot.lastKnownTargetId let targetId = if bot.currentTargetId >= 0: bot.currentTargetId else: bot.lastKnownTargetId
# Aggregate fitness across all tracked enemies (or just the primary target if known). # Per-target fitness when known, else a deterministic recency-weighted aggregate
# Use aggregate (targetId=-1) so we always have data even if target switched. # over all enemies (see VirtualTracker.fitnessFor). Using the shared proc keeps
let fit = block: # this stats dump identical to what the gun selector sees.
var f: seq[vb.GunFitness] let fit = bot.tracker.fitnessFor(targetId)
if targetId >= 0 and targetId in bot.tracker.fitness:
f = bot.tracker.fitness[targetId]
else:
# aggregate across all enemies
f = newSeq[vb.GunFitness](bot.tracker.numGuns)
for perEnemy in bot.tracker.fitness.values:
for gid in 0..<bot.tracker.numGuns:
for binIdx in 0..<len(vb.PowerBins):
let src = perEnemy[gid].bins[binIdx]
let n = min(src.count, vb.WindowSize)
for k in 0..<n:
f[gid].bins[binIdx].hits[(f[gid].bins[binIdx].head + f[gid].bins[binIdx].count) mod vb.WindowSize] = src.hits[k]
inc f[gid].bins[binIdx].count
f
var gunsArr = newJArray() var gunsArr = newJArray()
for gid in 0..<13: for gid in 0..<13:
@@ -251,7 +237,9 @@ method onRoundEnded*(bot: ModularBot, e: RoundEndedEventForBot) =
"realShots": bot.realShotsFired, "realShots": bot.realShotsFired,
"realHits": bot.realHits, "realHits": bot.realHits,
"score": e.results.totalScore, "score": e.results.totalScore,
"bulletDmg": e.results.bulletDamage "bulletDmg": e.results.bulletDamage,
"vDropped": bot.tracker.droppedBullets,
"vStarved": bot.guessFactor.waveStarved + bot.decayGF.waveStarved + bot.knnGun.waveStarved
} }
let f = open("/tmp/gun_stats.jsonl", fmAppend) let f = open("/tmp/gun_stats.jsonl", fmAppend)
f.writeLine($row) f.writeLine($row)
@@ -269,6 +257,15 @@ method onRoundStarted*(bot: ModularBot, e: RoundStartedEvent) =
bot.gunSelectionCount[i] = 0 bot.gunSelectionCount[i] = 0
bot.gunRealShots[i] = 0 bot.gunRealShots[i] = 0
bot.gunRealHits[i] = 0 bot.gunRealHits[i] = 0
# Reset per-round integrity counters so each /tmp/gun_stats.jsonl line reports
# that round's numbers (summing the lines gives the session total).
bot.tracker.droppedBullets = 0
bot.guessFactor.wavePushes = 0
bot.guessFactor.waveStarved = 0
bot.decayGF.wavePushes = 0
bot.decayGF.waveStarved = 0
bot.knnGun.wavePushes = 0
bot.knnGun.waveStarved = 0
setAdjustGunForBodyTurn(true) setAdjustGunForBodyTurn(true)
setAdjustRadarForBodyTurn(true) setAdjustRadarForBodyTurn(true)
setAdjustRadarForGunTurn(true) setAdjustRadarForGunTurn(true)
+27 -7
View File
@@ -6,12 +6,16 @@
import std/math import std/math
import std/tables import std/tables
import std/random import std/random
import std/algorithm
import gun_interface import gun_interface
const const
PowerBins* = [1.0, 1.5, 2.0, 3.0] ## 4 bins; ponytail: fixed array, add runtime config if needed PowerBins* = [1.0, 1.5, 2.0, 3.0] ## 4 bins; ponytail: fixed array, add runtime config if needed
WindowSize* = 100 ## rolling window ticks for fitness WindowSize* = 100 ## rolling window ticks for fitness
MaxBullets* = 2048 ## hard cap; ponytail: ring buffer, resize if more guns added MaxBullets* = 8192 ## hard cap; ring buffer. 52 spawns/tick and a
## full-map long shot (~90 ticks) need ~4700 slots;
## 8192 wraps only after ~157 ticks. Each VirtualBullet
## is 88 bytes, so this array costs ~704 KiB.
MinHitRate* = 0.40 ## 40% threshold for acceptable power selection MinHitRate* = 0.40 ## 40% threshold for acceptable power selection
MinObsBeforeCompete* = 50 ## min observations before a gun×bin enters competition MinObsBeforeCompete* = 50 ## min observations before a gun×bin enters competition
TieMargin* = 0.02 ## guns within this hit-rate margin of best are tied TieMargin* = 0.02 ## guns within this hit-rate margin of best are tied
@@ -45,6 +49,7 @@ type
head*: int ## ring buffer head head*: int ## ring buffer head
numGuns*: int numGuns*: int
fitness*: Table[int, seq[GunFitness]] ## keyed by enemy bot ID, indexed by GunId fitness*: Table[int, seq[GunFitness]] ## keyed by enemy bot ID, indexed by GunId
droppedBullets*: int ## unresolved bullets clobbered by the ring buffer (should stay 0)
proc initTracker*(numGuns: int): VirtualTracker = proc initTracker*(numGuns: int): VirtualTracker =
result.numGuns = numGuns result.numGuns = numGuns
@@ -75,6 +80,11 @@ proc spawnBullets*(t: var VirtualTracker, gunId: GunId,
let pred = predictions[binIdx] let pred = predictions[binIdx]
let fireDist = hypot(pred.x - state.selfX, pred.y - state.selfY) let fireDist = hypot(pred.x - state.selfX, pred.y - state.selfY)
let slot = t.head mod MaxBullets let slot = t.head mod MaxBullets
# Measurement integrity: if the slot we are about to overwrite still holds an
# unresolved bullet, that bullet will never be scored. Count it instead of
# silently dropping it (non-zero after a battle means MaxBullets is too small).
if t.bullets[slot].active:
inc t.droppedBullets
t.bullets[slot] = VirtualBullet( t.bullets[slot] = VirtualBullet(
gunId: gunId, gunId: gunId,
powerBin: binIdx, powerBin: binIdx,
@@ -140,20 +150,30 @@ proc tickBullets*(t: var VirtualTracker, state: WorldState,
onResolved(b.gunId, b.powerBin, fe) onResolved(b.gunId, b.powerBin, fe)
b.active = false b.active = false
proc fitnessFor(t: VirtualTracker, targetId: int): seq[GunFitness] = proc fitnessFor*(t: VirtualTracker, targetId: int): seq[GunFitness] =
## Returns fitness seq for targetId, or merges all enemies as fallback. ## Returns fitness seq for targetId, or merges all enemies as fallback.
## ponytail: merge is O(enemies*guns*bins), fine for small counts ##
## The fallback is a RECENCY-WEIGHTED AGGREGATE over the last WindowSize
## samples, NOT a pooled rate: each per-enemy window is replayed into one fresh
## window, so once the total exceeds WindowSize the earliest samples are
## overwritten by later ones. Enemies are visited in ascending target-id order
## so the result is identical on every run (std/tables iteration order is hash
## order and therefore nondeterministic).
## ponytail: merge is O(enemies*guns*bins*WindowSize), fine for small counts
if targetId >= 0 and targetId in t.fitness: if targetId >= 0 and targetId in t.fitness:
return t.fitness[targetId] return t.fitness[targetId]
# Aggregate across all enemies # Aggregate across all enemies, deterministically ordered.
result = newSeq[GunFitness](t.numGuns) result = newSeq[GunFitness](t.numGuns)
for perEnemy in t.fitness.values: var enemyIds: seq[int]
for id in t.fitness.keys: enemyIds.add id
enemyIds.sort()
for id in enemyIds:
let perEnemy = t.fitness[id]
for gunId in 0..<t.numGuns: for gunId in 0..<t.numGuns:
for binIdx in 0..<len(PowerBins): for binIdx in 0..<len(PowerBins):
let src = perEnemy[gunId].bins[binIdx] let src = perEnemy[gunId].bins[binIdx]
var dst = addr result[gunId].bins[binIdx]
for k in 0..<min(src.count, WindowSize): for k in 0..<min(src.count, WindowSize):
dst[].record(src.hits[k]) result[gunId].bins[binIdx].record(src.hits[k])
proc bestPower*(t: VirtualTracker, gunId: GunId, targetId: int = -1): (int, float) = proc bestPower*(t: VirtualTracker, gunId: GunId, targetId: int = -1): (int, float) =
## Returns (binIdx, power) with highest power that has >= MinHitRate. ## Returns (binIdx, power) with highest power that has >= MinHitRate.
+48 -11
View File
@@ -1,14 +1,16 @@
## Recency-weighted GF gun: exponential decay on histogram bins. ## Recency-weighted GF gun: exponential decay on histogram bins.
## decay=0.998/tick gives ~350-tick half-life — adapts to mid-battle strategy shifts. ## decay=0.998/tick gives ~350-tick half-life — adapts to mid-battle strategy shifts.
## Everything else identical to guess_factor.nim. ## Everything else identical to guess_factor.nim, including per-power-bin wave queues.
import std/math import std/math
import gun_harness/gun_interface import gun_harness/gun_interface
import gun_harness/virtual_bullets as vb # PowerBins
const const
GFBins = 31 GFBins = 31
GFPrior = 0.1 GFPrior = 0.1
DecayRate = 0.998 # ponytail: single global decay, tune if adaptation too slow/fast DecayRate = 0.998 # ponytail: single global decay, tune if adaptation too slow/fast
DecayWaveCompactAt = 64
type type
DWave = object DWave = object
@@ -17,14 +19,21 @@ type
DecayGFGun* = object DecayGFGun* = object
bins: array[GFBins, float] bins: array[GFBins, float]
waves: seq[DWave] # One wave queue per power bin; a resolved bullet only learns from a wave
cachedTick: int # queued for its own bin (matched on bulletSpeed / bulletPower).
cachedWaveStored: bool waves: array[len(vb.PowerBins), seq[DWave]]
waveHead: array[len(vb.PowerBins), int] # O(1) pop cursor
waveStoredTick: array[len(vb.PowerBins), int] # last tick a wave was queued for this bin
cachedTick: int # last tick bins were decayed
wavePushes*: int
waveStarved*: int
debugGraphics*: bool debugGraphics*: bool
proc initDecayGFGun*(): DecayGFGun = proc initDecayGFGun*(): DecayGFGun =
result.cachedTick = -1 result.cachedTick = -1
result.debugGraphics = false result.debugGraphics = false
for b in 0..<len(vb.PowerBins):
result.waveStoredTick[b] = -1
let center = (GFBins - 1) div 2 let center = (GFBins - 1) div 2
for i in 0..<GFBins: for i in 0..<GFBins:
let d = abs(i - center) let d = abs(i - center)
@@ -43,6 +52,28 @@ proc peakBin(g: DecayGFGun): int =
best = i best = i
best best
proc binForSpeed(spd: float): int {.inline.} =
for i in 0..<len(vb.PowerBins):
if abs(spd - bulletSpeed(vb.PowerBins[i])) < 1e-6:
return i
-1
proc binForPower(power: float): int {.inline.} =
for i in 0..<len(vb.PowerBins):
if abs(power - vb.PowerBins[i]) < 1e-6:
return i
-1
proc takeOldestWave(g: var DecayGFGun, binIdx: int): (bool, DWave) =
if binIdx < 0 or g.waveHead[binIdx] >= g.waves[binIdx].len:
return (false, DWave())
result = (true, g.waves[binIdx][g.waveHead[binIdx]])
inc g.waveHead[binIdx]
if g.waveHead[binIdx] >= DecayWaveCompactAt 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 predict*(g: var DecayGFGun, state: WorldState, bulletSpeed: float): GunPrediction = proc predict*(g: var DecayGFGun, state: WorldState, bulletSpeed: float): GunPrediction =
if bulletSpeed <= 0.0: if bulletSpeed <= 0.0:
return GunPrediction(x: state.enemyX, y: state.enemyY) return GunPrediction(x: state.enemyX, y: state.enemyY)
@@ -58,11 +89,14 @@ proc predict*(g: var DecayGFGun, state: WorldState, bulletSpeed: float): GunPred
for i in 0..<GFBins: for i in 0..<GFBins:
g.bins[i] *= DecayRate g.bins[i] *= DecayRate
g.cachedTick = state.tick g.cachedTick = state.tick
g.cachedWaveStored = false
if not g.cachedWaveStored: # Queue at most one wave per (tick, power bin); the fire site's extra predict()
g.waves.add DWave(fireX: state.selfX, fireY: state.selfY, fireBearing: bearing) # call for the selected bin lands on the same tick and reuses the queued wave.
g.cachedWaveStored = true let binIdx = binForSpeed(bulletSpeed)
if binIdx >= 0 and g.waveStoredTick[binIdx] != state.tick:
g.waves[binIdx].add DWave(fireX: state.selfX, fireY: state.selfY, fireBearing: bearing)
g.waveStoredTick[binIdx] = state.tick
inc g.wavePushes
let peak = g.peakBin() let peak = g.peakBin()
let peakGF = indexToGF(peak) let peakGF = indexToGF(peak)
@@ -76,10 +110,13 @@ proc predict*(g: var DecayGFGun, state: WorldState, bulletSpeed: float): GunPred
) )
proc onResult*(g: var DecayGFGun, e: FeedbackEvent) = proc onResult*(g: var DecayGFGun, e: FeedbackEvent) =
if g.waves.len == 0: return let binIdx = binForPower(e.bulletPower)
if binIdx < 0: return
let w = g.waves[0] let (found, w) = g.takeOldestWave(binIdx)
g.waves.delete(0) if not found:
inc g.waveStarved
return
let speed = bulletSpeed(e.bulletPower) let speed = bulletSpeed(e.bulletPower)
let mea = arcsin(clamp(8.0 / speed, -1.0, 1.0)) let mea = arcsin(clamp(8.0 / speed, -1.0, 1.0))
+57 -25
View File
@@ -1,14 +1,16 @@
## Guess-factor gun: statistical targeting via GF histogram. ## Guess-factor gun: statistical targeting via GF histogram.
## Bins: 31, ranging GF -1 (max CW escape) to +1 (max CCW escape). ## Bins: 31, ranging GF -1 (max CW escape) to +1 (max CCW escape).
## Learns from virtual bullet outcomes; caches wave state per-tick. ## Learns from virtual bullet outcomes; queues one wave per (tick, power bin).
import std/[math, strformat] import std/[math, strformat]
import gun_harness/gun_interface import gun_harness/gun_interface
import gun_harness/virtual_bullets as vb # PowerBins: the four power bins the harness spawns
const const
GFBins = 31 GFBins = 31
GFPrior = 0.1 GFPrior = 0.1
DebugGF* = false DebugGF* = false
WaveCompactAt = 64 ## compact a bin's wave seq once this many entries are consumed
type type
Wave = object Wave = object
@@ -18,15 +20,20 @@ type
GFGun* = object GFGun* = object
bins: array[GFBins, float] bins: array[GFBins, float]
waves: seq[Wave] # pending unresolved waves # One wave queue per power bin. The owning bin is fixed at push time (from the
# per-tick cache: store wave only once across multiple power-bin calls # bulletSpeed argument) and at pop time (from FeedbackEvent.bulletPower), so a
cachedTick: int # resolved bullet is always paired with a wave from its own bin.
cachedWaveStored: bool waves: array[len(vb.PowerBins), seq[Wave]]
waveHead: array[len(vb.PowerBins), int] # O(1) pop cursor into waves[bin]
waveStoredTick: array[len(vb.PowerBins), int] # last tick a wave was queued for this bin
wavePushes*: int # total waves enqueued (== one per (tick, bin))
waveStarved*: int # onResult found an empty queue for its own bin
debugGraphics*: bool debugGraphics*: bool
proc initGFGun*(): GFGun = proc initGFGun*(): GFGun =
result.cachedTick = -1
result.debugGraphics = false result.debugGraphics = false
for b in 0..<len(vb.PowerBins):
result.waveStoredTick[b] = -1
# Seed with a head-on prior: triangular bump at bin 15 (GF=0). # Seed with a head-on prior: triangular bump at bin 15 (GF=0).
# Prevents the cold-start tie-break to GF=-1 (bin 0) that poisons early fitness. # Prevents the cold-start tie-break to GF=-1 (bin 0) that poisons early fitness.
let center = (GFBins - 1) div 2 # = 15 let center = (GFBins - 1) div 2 # = 15
@@ -47,6 +54,34 @@ proc peakBin(g: GFGun): int =
best = i best = i
best best
proc binForSpeed(spd: float): int {.inline.} =
## Map a virtual-bullet speed back to its power-bin index. All four bin speeds
## are exactly representable floats; the epsilon is belt-and-braces only.
for i in 0..<len(vb.PowerBins):
if abs(spd - bulletSpeed(vb.PowerBins[i])) < 1e-6:
return i
-1
proc binForPower(power: float): int {.inline.} =
## Map a FeedbackEvent.bulletPower back to its power-bin index.
for i in 0..<len(vb.PowerBins):
if abs(power - vb.PowerBins[i]) < 1e-6:
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 predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPrediction = proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPrediction =
if bulletSpeed <= 0.0: if bulletSpeed <= 0.0:
return GunPrediction(x: state.enemyX, y: state.enemyY) return GunPrediction(x: state.enemyX, y: state.enemyY)
@@ -57,18 +92,17 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
let bearing = arctan2(dy, dx) let bearing = arctan2(dy, dx)
let mea = arcsin(clamp(8.0 / bulletSpeed, -1.0, 1.0)) let mea = arcsin(clamp(8.0 / bulletSpeed, -1.0, 1.0))
# Store one wave per tick regardless of how many power bins call us # Queue at most one wave per (tick, power bin). The fire site's extra predict()
if state.tick != g.cachedTick: # call for the selected bin lands on the same tick and reuses the queued wave.
g.cachedTick = state.tick let binIdx = binForSpeed(bulletSpeed)
g.cachedWaveStored = false if binIdx >= 0 and g.waveStoredTick[binIdx] != state.tick:
g.waves[binIdx].add Wave(
if not g.cachedWaveStored:
g.waves.add Wave(
fireX: state.selfX, fireX: state.selfX,
fireY: state.selfY, fireY: state.selfY,
fireBearing: bearing, fireBearing: bearing,
) )
g.cachedWaveStored = true g.waveStoredTick[binIdx] = state.tick
inc g.wavePushes
let peak = g.peakBin() let peak = g.peakBin()
let peakGF = indexToGF(peak) let peakGF = indexToGF(peak)
@@ -78,7 +112,7 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
let py = state.selfY + sin(gfAngle) * dist let py = state.selfY + sin(gfAngle) * dist
when DebugGF: when DebugGF:
echo fmt"[gf-dbg] predict: peakGF={peakGF:.2f} peakBin={peak} mea={radToDeg(mea):.1f}° aimAngle={radToDeg(gfAngle):.1f}° waves={g.waves.len}" echo fmt"[gf-dbg] predict: peakGF={peakGF:.2f} peakBin={peak} mea={radToDeg(mea):.1f}° aimAngle={radToDeg(gfAngle):.1f}° waves={g.waves[binIdx].len}"
GunPrediction( GunPrediction(
x: clamp(px, BotRadius, state.arenaWidth - BotRadius), x: clamp(px, BotRadius, state.arenaWidth - BotRadius),
@@ -86,17 +120,15 @@ proc predict*(g: var GFGun, state: WorldState, bulletSpeed: float): GunPredictio
) )
proc onResult*(g: var GFGun, e: FeedbackEvent) = proc onResult*(g: var GFGun, e: FeedbackEvent) =
## Called when a virtual bullet resolves. Match the wave by predicted point, ## Called when a virtual bullet resolves. Pop the OLDEST unresolved wave from
## compute actual GF, and increment the histogram with a smoothing kernel. ## this bullet's own power-bin queue, compute the actual GF, and smooth-add it.
## We don't have the original wave tick here, so we use the prediction coords let binIdx = binForPower(e.bulletPower)
## to identify and remove the matching wave. if binIdx < 0: return
## ponytail: O(n) scan over waves; waves list stays tiny (< a dozen at a time)
if g.waves.len == 0:
return
# Pop the oldest wave (FIFO matches bullet resolution order) let (found, w) = g.takeOldestWave(binIdx)
let w = g.waves[0] if not found:
g.waves.delete(0) inc g.waveStarved
return
# Recompute mea from the actual bullet power (correct per-bin, not the cached first-bin mea) # Recompute mea from the actual bullet power (correct per-bin, not the cached first-bin mea)
let speed = bulletSpeed(e.bulletPower) let speed = bulletSpeed(e.bulletPower)
+55 -11
View File
@@ -5,6 +5,7 @@
import std/[math] import std/[math]
import gun_harness/gun_interface import gun_harness/gun_interface
import gun_harness/virtual_bullets as vb # PowerBins
const const
MaxObs = 2000 # ring-buffer cap MaxObs = 2000 # ring-buffer cap
@@ -25,10 +26,14 @@ type
KNNGun* = object KNNGun* = object
obs: seq[Obs] obs: seq[Obs]
obsHead: int # ring-buffer write index obsHead: int # ring-buffer write index
waves: seq[KNNWave] # One wave queue per power bin; matched on bulletSpeed / bulletPower so a
# resolved bullet only ever learns from a wave fired with the same power.
waves: array[len(vb.PowerBins), seq[KNNWave]]
waveHead: array[len(vb.PowerBins), int] # O(1) pop cursor
waveStoredTick: array[len(vb.PowerBins), int] # last tick a wave was queued for this bin
# per-tick cache # per-tick cache
cachedTick: int cachedTick: int
cachedWaveStored: bool tickWave: KNNWave # wave template for the current tick (features computed once)
# rolling normalization ranges # rolling normalization ranges
featMin: array[7, float] featMin: array[7, float]
featMax: array[7, float] featMax: array[7, float]
@@ -36,12 +41,16 @@ type
lastSpeed: float lastSpeed: float
lastDirection: float # +1 or -1 lastDirection: float # +1 or -1
timeSinceDirChange: int timeSinceDirChange: int
wavePushes*: int
waveStarved*: int
debugGraphics*: bool debugGraphics*: bool
proc initKNNGun*(): KNNGun = proc initKNNGun*(): KNNGun =
result.cachedTick = -1 result.cachedTick = -1
result.lastDirection = 1.0 result.lastDirection = 1.0
result.debugGraphics = false result.debugGraphics = false
for b in 0..<len(vb.PowerBins):
result.waveStoredTick[b] = -1
for i in 0..6: for i in 0..6:
result.featMin[i] = 1e18 result.featMin[i] = 1e18
result.featMax[i] = -1e18 result.featMax[i] = -1e18
@@ -105,6 +114,32 @@ proc euclidean(a, b: array[7, float]): float {.inline.} =
result += d * d result += d * d
result = sqrt(result) result = sqrt(result)
proc binForSpeed(spd: float): int {.inline.} =
## Map a virtual-bullet speed back to its power-bin index. All four bin speeds
## are exactly representable floats; the epsilon is belt-and-braces only.
for i in 0..<len(vb.PowerBins):
if abs(spd - bulletSpeed(vb.PowerBins[i])) < 1e-6:
return i
-1
proc binForPower(power: float): int {.inline.} =
## Map a FeedbackEvent.bulletPower back to its power-bin index.
for i in 0..<len(vb.PowerBins):
if abs(power - vb.PowerBins[i]) < 1e-6:
return i
-1
proc takeOldestWave(g: var KNNGun, binIdx: int): (bool, KNNWave) =
## Pop the oldest unresolved wave for this bin (O(1) amortized via waveHead).
if binIdx < 0 or g.waveHead[binIdx] >= g.waves[binIdx].len:
return (false, KNNWave())
result = (true, g.waves[binIdx][g.waveHead[binIdx]])
inc g.waveHead[binIdx]
if g.waveHead[binIdx] >= 64 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
# ── Gun interface ───────────────────────────────────────────────────────────── # ── Gun interface ─────────────────────────────────────────────────────────────
proc predict*(g: var KNNGun, state: WorldState, bulletSpd: float): GunPrediction = proc predict*(g: var KNNGun, state: WorldState, bulletSpd: float): GunPrediction =
@@ -120,7 +155,6 @@ proc predict*(g: var KNNGun, state: WorldState, bulletSpd: float): GunPrediction
# Track direction change — update state once per tick # Track direction change — update state once per tick
if state.tick != g.cachedTick: if state.tick != g.cachedTick:
g.cachedTick = state.tick g.cachedTick = state.tick
g.cachedWaveStored = false
let relHead = state.enemyHeading - bearing let relHead = state.enemyHeading - bearing
let latVel = state.enemySpeed * sin(relHead) let latVel = state.enemySpeed * sin(relHead)
@@ -131,18 +165,26 @@ proc predict*(g: var KNNGun, state: WorldState, bulletSpd: float): GunPrediction
else: else:
inc g.timeSinceDirChange inc g.timeSinceDirChange
# Store wave once per tick # Compute the tick's feature vector ONCE, before lastSpeed is advanced, so
if not g.cachedWaveStored: # every power bin fired this tick shares identical features. lastSpeed is
# only advanced here (once/tick), not once per bin.
let feat = buildFeatures(state, g.lastSpeed, g.lastDirection, g.timeSinceDirChange) let feat = buildFeatures(state, g.lastSpeed, g.lastDirection, g.timeSinceDirChange)
g.updateMinMax(feat) g.updateMinMax(feat)
g.waves.add KNNWave( g.tickWave = KNNWave(
fireX: state.selfX, fireX: state.selfX,
fireY: state.selfY, fireY: state.selfY,
fireBearing: bearing, fireBearing: bearing,
feat: feat, feat: feat,
) )
g.lastSpeed = state.enemySpeed g.lastSpeed = state.enemySpeed
g.cachedWaveStored = true
# Queue at most one wave per (tick, power bin). The fire site's extra predict()
# call for the selected bin lands on the same tick and reuses the queued wave.
let binIdx = binForSpeed(bulletSpd)
if binIdx >= 0 and g.waveStoredTick[binIdx] != state.tick:
g.waves[binIdx].add g.tickWave
g.waveStoredTick[binIdx] = state.tick
inc g.wavePushes
# Cold start — no data yet # Cold start — no data yet
if g.obs.len == 0: if g.obs.len == 0:
@@ -236,10 +278,12 @@ proc predict*(g: var KNNGun, state: WorldState, bulletSpd: float): GunPrediction
) )
proc onResult*(g: var KNNGun, e: FeedbackEvent) = proc onResult*(g: var KNNGun, e: FeedbackEvent) =
if g.waves.len == 0: return let binIdx = binForPower(e.bulletPower)
if binIdx < 0: return
let w = g.waves[0] let (found, w) = g.takeOldestWave(binIdx)
g.waves.delete(0) if not found:
inc g.waveStarved
return
let speed = bulletSpeed(e.bulletPower) let speed = bulletSpeed(e.bulletPower)
let mea = arcsin(clamp(8.0 / speed, -1.0, 1.0)) let mea = arcsin(clamp(8.0 / speed, -1.0, 1.0))