SirStone
0d35646dc9
- actorForward: deterministic param, uses mean-only when PPOB_EVAL_ONLY=1 (eval was adding unit Gaussian noise to every action — unreliable scores) - warm_start.py: log_std initialized to -1.0 (std≈0.37) instead of copying snapshot values (were 2.27-4.68 → std 9-108, completely drowning signal) - training.env: LOG_STD_CEILING 0.0→-0.5 (cap exploration at std≈0.6)
372 lines
17 KiB
Nim
372 lines
17 KiB
Nim
## PPO_Bot — enemy tracker + state vector wired into the game loop.
|
|
## Training: trajectory collected per tick, PPO update in background thread.
|
|
|
|
import std/[os, strformat, strutils, math, times, algorithm]
|
|
import arraymancer
|
|
import tankroyale_botapi
|
|
import network
|
|
import actions
|
|
import training
|
|
import weights
|
|
import ./enemy_tracker
|
|
import ./state_vector
|
|
|
|
# ── Hyperparameters from env vars (PPOB_ prefix) ─────────────────────────────
|
|
# All optional; defaults match ppoUpdate signature in training.nim.
|
|
|
|
proc getEnvFloat(name: string, default: float32): float32 =
|
|
let v = getEnv(name)
|
|
if v.len == 0: default else: parseFloat(v).float32
|
|
|
|
proc getEnvInt(name: string, default: int): int =
|
|
let v = getEnv(name)
|
|
if v.len == 0: default else: parseInt(v)
|
|
|
|
var
|
|
hpLr: float32 = getEnvFloat("PPOB_LR", 3e-4'f32)
|
|
hpClipEpsilon: float32 = getEnvFloat("PPOB_CLIP_EPSILON", 0.2'f32)
|
|
hpEntropyCoeff: float32 = getEnvFloat("PPOB_ENTROPY_COEFF", 0.01'f32)
|
|
hpValueLossCoeff: float32 = getEnvFloat("PPOB_VALUE_LOSS_COEFF", 0.5'f32)
|
|
hpMaxGradNorm: float32 = getEnvFloat("PPOB_MAX_GRAD_NORM", 0.5'f32)
|
|
hpGamma: float32 = getEnvFloat("PPOB_GAMMA", 0.99'f32)
|
|
hpLam: float32 = getEnvFloat("PPOB_LAM", 0.95'f32)
|
|
hpEpochs: int = getEnvInt("PPOB_EPOCHS", 4)
|
|
hpMiniBatchSize: int = getEnvInt("PPOB_MINI_BATCH_SIZE", 64)
|
|
hpUpdateInterval: int = getEnvInt("PPOB_UPDATE_INTERVAL", 10)
|
|
|
|
# Wire logStd tunable params into network module vars (read before initActorCritic)
|
|
logStdFloor = getEnvFloat("PPOB_LOG_STD_FLOOR", -3.0'f32)
|
|
logStdCeiling = getEnvFloat("PPOB_LOG_STD_CEILING", 0.5'f32)
|
|
initialLogStd = getEnvFloat("PPOB_INITIAL_LOG_STD", 0.0'f32)
|
|
|
|
# ── Structured log output ─────────────────────────────────────────────────────
|
|
|
|
let logFile = getEnv("PPOB_LOG_FILE") # empty → no JSON logging
|
|
let evalOnly = getEnv("PPOB_EVAL_ONLY") == "1" # freeze training (pure evaluation)
|
|
|
|
proc appendJsonLine(path, line: string) =
|
|
## Append a JSON line to path; no-op if path is empty.
|
|
if path.len == 0: return
|
|
let f = open(path, fmAppend)
|
|
f.writeLine(line)
|
|
f.close()
|
|
|
|
proc jsonFloat(v: float32): string =
|
|
## Serialize a float for JSON; non-finite → null (keeps JSONL parseable).
|
|
if v == v and v > -1e30'f32 and v < 1e30'f32: $v else: "null"
|
|
|
|
proc hyperparmSnapshot(): string =
|
|
## Compact JSON object of current hyperparams (no outer braces).
|
|
&"\"lr\":{hpLr},\"clipEpsilon\":{hpClipEpsilon}," &
|
|
&"\"entropyCoeff\":{hpEntropyCoeff},\"valueLossCoeff\":{hpValueLossCoeff}," &
|
|
&"\"maxGradNorm\":{hpMaxGradNorm},\"gamma\":{hpGamma},\"lam\":{hpLam}," &
|
|
&"\"epochs\":{hpEpochs},\"miniBatchSize\":{hpMiniBatchSize}," &
|
|
&"\"logStdFloor\":{logStdFloor},\"logStdCeiling\":{logStdCeiling},\"initialLogStd\":{initialLogStd}"
|
|
|
|
const botJsonPath = currentSourcePath().parentDir / "PPO_Bot.json"
|
|
const weightsRoot = currentSourcePath().parentDir / "weights"
|
|
|
|
type
|
|
InFlightBullet = object
|
|
x, y: float64 # current position
|
|
vx, vy: float64 # velocity (pixels/tick)
|
|
power: float64
|
|
|
|
type PPOBot = ref object of Bot
|
|
tracker: EnemyTracker
|
|
buffer: TrajectoryBuffer
|
|
prevEnergy: float32 # own energy last tick
|
|
prevEnemyE: float32 # enemy energy last tick (from tracker)
|
|
lastState: array[STATE_DIM, float32] # plain arrays — tensors NEVER cross threads
|
|
lastAction: array[ACTION_DIM, float32]
|
|
lastLogP: float32
|
|
lastValue: float32
|
|
hasLastTrans: bool
|
|
lastActions: BotActions # previous tick's decoded actions (for state vector)
|
|
roundRewardSum: float32 # cumulative reward this round (for live display)
|
|
roundTicks: int # ticks this round
|
|
# Fixed-size bullet buffer — NO heap on the shared bot object. A seq here is
|
|
# allocated by the per-round bot thread and freed by the next round's thread
|
|
# (bot.bullets = @[] on round start) → foreign-heap free under --threads:on +
|
|
# ORC → SIGSEGV. N=4: the state vector only consumes the closest 3 slots.
|
|
bullets: array[4, InFlightBullet]
|
|
bulletCount: int
|
|
|
|
var ac = initActorCritic()
|
|
var gAdamStates: ACAdamStates # persists across rounds
|
|
var roundCounter = 0
|
|
var roundsSinceUpdate = 0
|
|
|
|
# ── Bot methods ───────────────────────────────────────────────────────────────
|
|
|
|
method onScannedBot*(bot: PPOBot, e: ScannedBotEvent) =
|
|
bot.tracker.update(e.x, e.y, e.direction, e.speed, e.energy)
|
|
|
|
method onRoundStarted*(bot: PPOBot, e: RoundStartedEvent) =
|
|
setAdjustGunForBodyTurn(true)
|
|
setAdjustRadarForBodyTurn(true)
|
|
setAdjustRadarForGunTurn(true)
|
|
bot.tracker = initEnemyTracker()
|
|
# Do NOT clear bot.buffer here — transitions accumulate across rounds
|
|
# until hpUpdateInterval rounds have passed (cleared in onRoundEnded).
|
|
bot.prevEnergy = 0.0'f32
|
|
bot.prevEnemyE = 0.0'f32
|
|
bot.hasLastTrans = false
|
|
bot.roundRewardSum = 0.0'f32
|
|
bot.roundTicks = 0
|
|
bot.bulletCount = 0
|
|
|
|
method onRoundEnded*(bot: PPOBot, e: RoundEndedEventForBot) =
|
|
inc roundCounter
|
|
inc roundsSinceUpdate
|
|
debugLog("[PO-ENTER] round=" & $roundCounter & " tid=" & $getThreadId())
|
|
|
|
# Add round-end score bonus to last transition and mark it as episode boundary
|
|
let roundReward = computeRoundReward(e.results.totalScore.float32)
|
|
if bot.hasLastTrans and bot.buffer.len > 0:
|
|
let lastIdx = bot.buffer.len - 1
|
|
bot.buffer.transitions[lastIdx].reward += roundReward
|
|
bot.buffer.transitions[lastIdx].done = true
|
|
|
|
# Training progress display — one line per round in the UI console
|
|
let ticks = bot.buffer.len
|
|
var avgR = 0.0'f32
|
|
if ticks > 0:
|
|
var rewardSum = 0.0'f32
|
|
for i in 0 ..< bot.buffer.len: rewardSum += bot.buffer.transitions[i].reward
|
|
avgR = rewardSum / ticks.float32
|
|
let avgRStr = formatFloat(avgR.float, ffDecimal, 3)
|
|
printToStdOut(&"R:{roundCounter} ticks:{ticks} buf:{bot.buffer.len} sinceUpd:{roundsSinceUpdate}/{hpUpdateInterval} avgR:{avgRStr} score:{e.results.totalScore}\n")
|
|
echo &"R:{roundCounter} ticks:{ticks} buf:{bot.buffer.len} sinceUpd:{roundsSinceUpdate}/{hpUpdateInterval} avgR:{avgRStr} score:{e.results.totalScore}"
|
|
|
|
writeFile(weightsRoot / "round_counter.txt", $roundCounter)
|
|
|
|
if bot.buffer.len == 0:
|
|
bot.hasLastTrans = false
|
|
return
|
|
|
|
# Emit per-round game-stats JSON line
|
|
let ts = int(epochTime())
|
|
let jline = &"""{{"type":"round","round":{roundCounter},"ticks":{ticks},"bufLen":{bot.buffer.len},"avgReward":{avgR},"score":{e.results.totalScore},"ts":{ts}}}"""
|
|
appendJsonLine(logFile, jline)
|
|
|
|
# PPOB_EVAL_ONLY=1 → freeze training (pure evaluation): skip ppoUpdate and
|
|
# checkpoint save, but keep advancing/writing round_counter.txt so run.sh's
|
|
# remaining-rounds bookkeeping still works, and keep the game line above.
|
|
if evalOnly:
|
|
bot.buffer.clear()
|
|
bot.hasLastTrans = false
|
|
roundsSinceUpdate = 0
|
|
return
|
|
|
|
# Always save weights every round so round_counter.txt stays current.
|
|
saveCheckpoint(ac, gAdamStates, weightsRoot, roundCounter)
|
|
|
|
# Only update policy every hpUpdateInterval rounds (~3000 transitions).
|
|
if roundsSinceUpdate >= hpUpdateInterval:
|
|
# ponytail: synchronous update. Arraymancer tensors can't cross threads under
|
|
# ORC — training-thread ppoUpdate frees/rebinds tensors owned by the bot
|
|
# thread's heap (SIGSEGV, reproduced with a lone trainer thread on a fixed
|
|
# buffer; save/channel/forward exonerated). The bot API runs events on one
|
|
# bot thread, so inline is single-threaded. Revert to a background thread
|
|
# only if tensors are rebuilt from plain data on that thread.
|
|
printToStdOut(&" train→ R:{roundCounter} bufLen:{bot.buffer.len}\n")
|
|
echo &" train→ R:{roundCounter} bufLen:{bot.buffer.len}"
|
|
let m = ppoUpdate(ac, bot.buffer,
|
|
lastValue = 0.0'f32,
|
|
adamStates = gAdamStates,
|
|
epochs = hpEpochs,
|
|
miniBatchSize = hpMiniBatchSize,
|
|
clipEpsilon = hpClipEpsilon,
|
|
entropyCoeff = hpEntropyCoeff,
|
|
valueLossCoeff = hpValueLossCoeff,
|
|
lr = hpLr,
|
|
maxGradNorm = hpMaxGradNorm,
|
|
gamma = hpGamma,
|
|
lam = hpLam)
|
|
printToStdOut(&" trained R:{roundCounter} aLoss:{formatFloat(m.actorLoss.float, ffDecimal, 4)} vLoss:{formatFloat(m.valueLoss.float, ffDecimal, 4)} gNorm:{formatFloat(m.gradNorm.float, ffDecimal, 3)}\n")
|
|
echo &" trained R:{roundCounter} aLoss:{formatFloat(m.actorLoss.float, ffDecimal, 4)} vLoss:{formatFloat(m.valueLoss.float, ffDecimal, 4)} gNorm:{formatFloat(m.gradNorm.float, ffDecimal, 3)}"
|
|
# Emit training-health JSON line
|
|
let hp = hyperparmSnapshot()
|
|
let ts2 = int(epochTime())
|
|
let jline2 = &"""{{"type":"train","round":{roundCounter},"actorLoss":{jsonFloat(m.actorLoss)},"valueLoss":{jsonFloat(m.valueLoss)},"gradNorm":{jsonFloat(m.gradNorm)},"ts":{ts2},{hp}}}"""
|
|
appendJsonLine(logFile, jline2)
|
|
bot.buffer.clear()
|
|
roundsSinceUpdate = 0
|
|
|
|
bot.hasLastTrans = false
|
|
debugLog("[PO-EXIT] round=" & $roundCounter & " tid=" & $getThreadId())
|
|
|
|
method run(bot: PPOBot) =
|
|
debugLog("[RUN-ENTER] tid=" & $getThreadId())
|
|
# Seed energy and goto/aimTo targets on first tick (remainingDistance = 0 initially)
|
|
bot.prevEnergy = getEnergy().float32
|
|
bot.prevEnemyE = if bot.tracker.hasContact: bot.tracker.current.energy.float32 else: 0.0'f32
|
|
bot.lastActions.gotoX = getX()
|
|
bot.lastActions.gotoY = getY()
|
|
bot.lastActions.aimToX = getX()
|
|
bot.lastActions.aimToY = getY()
|
|
|
|
while isRunning():
|
|
bot.tracker.deadReckon()
|
|
|
|
# Spawn bullet when enemy fired last scan tick. Edge-triggered: the flag is
|
|
# consumed later (after the state build) so a radar gap (deadReckon ticks)
|
|
# can't spawn k phantom bullets from one shot — exactly one bullet, and
|
|
# state index 12 still pulses 1 on the detection tick.
|
|
if bot.tracker.hasContact and bot.tracker.current.hasFired:
|
|
if bot.bulletCount < 4:
|
|
let power = bot.tracker.current.lastFirePower
|
|
let speed = 20.0 - 3.0 * power
|
|
# Approximate gun direction: bearing from enemy toward our position
|
|
let myX = getX(); let myY = getY()
|
|
let ang = arctan2(myY - bot.tracker.current.y, myX - bot.tracker.current.x)
|
|
bot.bullets[bot.bulletCount] = InFlightBullet(
|
|
x: bot.tracker.current.x,
|
|
y: bot.tracker.current.y,
|
|
vx: speed * cos(ang),
|
|
vy: speed * sin(ang),
|
|
power: power,
|
|
)
|
|
inc bot.bulletCount
|
|
|
|
# Advance in-flight bullets and prune those off-arena (in-place compaction
|
|
# into the fixed buffer — no per-tick heap churn).
|
|
let aW = float64(getArenaWidth()); let aH = float64(getArenaHeight())
|
|
var n = 0
|
|
for i in 0 ..< bot.bulletCount:
|
|
let b = bot.bullets[i]
|
|
let nx = b.x + b.vx; let ny = b.y + b.vy
|
|
if nx >= 0.0 and nx <= aW and ny >= 0.0 and ny <= aH:
|
|
bot.bullets[n] = InFlightBullet(x: nx, y: ny, vx: b.vx, vy: b.vy, power: b.power)
|
|
inc n
|
|
bot.bulletCount = n
|
|
|
|
# Convert to BulletData for state vector (closest 3 by distance to us)
|
|
let myX2 = getX(); let myY2 = getY()
|
|
var bulletData: array[4, BulletData]
|
|
var bulletDataCount = 0
|
|
for i in 0 ..< bot.bulletCount:
|
|
let b = bot.bullets[i]
|
|
bulletData[bulletDataCount] = BulletData(x: b.x, y: b.y, power: b.power)
|
|
inc bulletDataCount
|
|
# sort ascending by distance so the nearest threats fill slots 0-2
|
|
if bulletDataCount > 1:
|
|
bulletData.toOpenArray(0, bulletDataCount - 1).sort(proc(a, b: BulletData): int =
|
|
let da = hypot(a.x - myX2, a.y - myY2)
|
|
let db = hypot(b.x - myX2, b.y - myY2)
|
|
cmp(da, db))
|
|
|
|
setRadarTurnRate(bot.tracker.getRadarTurnRate(getX(), getY(), getDirection(), getRadarDirection()))
|
|
|
|
let botData = BotStateData(
|
|
x: getX(),
|
|
y: getY(),
|
|
direction: getDirection(),
|
|
speed: getSpeed(),
|
|
energy: getEnergy(),
|
|
gunDirection: getGunDirection(),
|
|
gunHeat: getGunHeat(),
|
|
arenaWidth: float64(getArenaWidth()),
|
|
arenaHeight: float64(getArenaHeight()),
|
|
)
|
|
|
|
let remainingGotoDistance = hypot(bot.lastActions.gotoX - botData.x,
|
|
bot.lastActions.gotoY - botData.y)
|
|
let remainingGunAngle = abs(normalizeRelativeAngle(
|
|
directionTo(botData.x, botData.y, bot.lastActions.aimToX, bot.lastActions.aimToY) -
|
|
botData.gunDirection))
|
|
let state = buildStateVector(botData, bot.tracker, remainingGotoDistance, remainingGunAngle, bulletData, bulletDataCount)
|
|
# Consume the fired flag AFTER the state build: state index 12 saw the
|
|
# detection-tick pulse, and the next iteration's spawn check sees false —
|
|
# one shot → exactly one bullet, even across deadReckon gaps.
|
|
bot.tracker.current.hasFired = false
|
|
let (rawActs, logP) = ac.actorForward(state, deterministic = evalOnly)
|
|
let value = ac.criticForward(state)
|
|
let ex = if bot.tracker.hasContact: bot.tracker.current.x else: botData.arenaWidth / 2.0
|
|
let ey = if bot.tracker.hasContact: bot.tracker.current.y else: botData.arenaHeight / 2.0
|
|
let acts = mapActions(rawActs,
|
|
getGunHeat().float,
|
|
botData.arenaWidth, botData.arenaHeight,
|
|
botData.x, botData.y,
|
|
botData.direction, botData.speed, botData.gunDirection,
|
|
ex, ey)
|
|
|
|
# Compute tick reward from energy deltas + dense shaping
|
|
let curEnergy = getEnergy().float32
|
|
let curEnemyE = if bot.tracker.hasContact: bot.tracker.current.energy.float32 else: bot.prevEnemyE
|
|
let myDelta = curEnergy - bot.prevEnergy
|
|
let enemyDelta = curEnemyE - bot.prevEnemyE
|
|
let arenaDiag = float32(sqrt(botData.arenaWidth * botData.arenaWidth +
|
|
botData.arenaHeight * botData.arenaHeight))
|
|
let distEnemy = if bot.tracker.hasContact:
|
|
float32(hypot(bot.tracker.current.x - botData.x,
|
|
bot.tracker.current.y - botData.y))
|
|
else: arenaDiag
|
|
let gunToEnemy = if bot.tracker.hasContact:
|
|
abs(normalizeRelativeAngle(
|
|
arctan2(bot.tracker.current.y - botData.y,
|
|
bot.tracker.current.x - botData.x) * 180.0 / PI -
|
|
botData.gunDirection)).float32
|
|
else: 180.0'f32
|
|
let tickReward = computeTickReward(myDelta, enemyDelta,
|
|
distToEnemy = distEnemy,
|
|
maxDist = arenaDiag,
|
|
gunBearingAbs = gunToEnemy)
|
|
|
|
# Track running reward for in-game display
|
|
bot.roundRewardSum += tickReward
|
|
inc bot.roundTicks
|
|
|
|
# Finalise previous transition with the reward from this tick's state change
|
|
if bot.hasLastTrans:
|
|
let tr = Transition(
|
|
state: bot.lastState,
|
|
action: bot.lastAction,
|
|
logProb: bot.lastLogP,
|
|
reward: tickReward,
|
|
value: bot.lastValue,
|
|
done: false, # episode boundary set in onRoundEnded
|
|
)
|
|
bot.buffer.add(tr)
|
|
|
|
# Store current for next tick — plain arrays only. `state`/`rawActs` tensors
|
|
# live and die on this thread; a NEW bot thread runs each round, so storing
|
|
# tensors in the shared bot object would free round-N's heap memory from
|
|
# round N+1's thread (SIGSEGV; confirmed empirically).
|
|
bot.lastState = stateToArr(state)
|
|
bot.lastAction = actionToArr(rawActs)
|
|
bot.lastLogP = logP
|
|
bot.lastValue = value
|
|
bot.prevEnergy = curEnergy
|
|
bot.prevEnemyE = curEnemyE
|
|
bot.hasLastTrans = true
|
|
bot.lastActions = acts
|
|
|
|
setTargetSpeed(acts.targetSpeed.float)
|
|
setTurnRate(acts.turnRate.float)
|
|
setGunTurnRate(acts.gunTurnRate.float)
|
|
if acts.shouldFire:
|
|
discard setFire(acts.firePower.float)
|
|
|
|
# In-game training progress overlay
|
|
let avgR = if bot.roundTicks > 0: bot.roundRewardSum / bot.roundTicks.float32
|
|
else: 0.0'f32
|
|
let avgRStr = formatFloat(avgR.float, ffDecimal, 2)
|
|
drawText(&"R:{roundCounter} avg:{avgRStr}", getX(), getY() - 40.0)
|
|
|
|
go()
|
|
|
|
when isMainModule:
|
|
createDir(weightsRoot)
|
|
cleanStaleTempDirs(weightsRoot)
|
|
let loadResult = loadBestAvailable(ac, gAdamStates, weightsRoot)
|
|
if loadResult.loaded:
|
|
roundCounter = loadResult.roundNum
|
|
|
|
var bot = PPOBot(
|
|
tracker: initEnemyTracker(),
|
|
buffer: initTrajectoryBuffer(),
|
|
)
|
|
start(bot, botJsonPath)
|