SirStone
766b9e03ee
- actions.nim: goto/aimTo coordinates now offset from enemy position (enemyX + tanh(raw) * scale) instead of absolute arena coords (sigmoid(raw) * arenaSize). Initial random policy defaults to approaching and aiming at enemy. - training.nim: added dense reward shaping (distance closeness + gun bearing) to computeTickReward, doubled round reward scaling. - PPO_Bot.nim: passes enemy position to mapActions, computes gun-to-enemy bearing for reward shaping. Result: 100/100 win rate vs Target with frozen weights. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
339 lines
14 KiB
Nim
339 lines
14 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]
|
|
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)
|
|
|
|
# Wire logStd tunable params into network module vars (read before initActorCritic)
|
|
logStdFloor = getEnvFloat("PPOB_LOG_STD_FLOOR", -3.0'f32)
|
|
initialLogStd = getEnvFloat("PPOB_INITIAL_LOG_STD", 0.0'f32)
|
|
|
|
# ── Structured log output ─────────────────────────────────────────────────────
|
|
|
|
let logFile = getEnv("PPOB_LOG_FILE") # empty → no JSON logging
|
|
|
|
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 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},\"initialLogStd\":{initialLogStd}"
|
|
|
|
const botJsonPath = currentSourcePath().parentDir / "PPO_Bot.json"
|
|
const weightsRoot = currentSourcePath().parentDir / "weights"
|
|
|
|
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: Tensor[float32]
|
|
lastAction: Tensor[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
|
|
|
|
var ac = initActorCritic()
|
|
var gAdamStates: ACAdamStates # persists across rounds
|
|
|
|
# ── Background training state ─────────────────────────────────────────────────
|
|
|
|
type
|
|
TrainingResult = object
|
|
ac: ActorCritic
|
|
adamStates: ACAdamStates
|
|
metrics: PPOMetrics
|
|
|
|
TrainingArgs = object
|
|
ac: ActorCritic
|
|
adamStates: ACAdamStates
|
|
buffer: TrajectoryBuffer
|
|
lastValue: float32
|
|
roundNum: int
|
|
weightsRoot: string
|
|
# hyperparams snapshot at launch time
|
|
lr: float32
|
|
clipEpsilon: float32
|
|
entropyCoeff: float32
|
|
valueLossCoeff: float32
|
|
maxGradNorm: float32
|
|
gamma: float32
|
|
lam: float32
|
|
epochs: int
|
|
miniBatchSize: int
|
|
|
|
var
|
|
trainingThread: Thread[TrainingArgs]
|
|
resultChan: Channel[TrainingResult]
|
|
threadLaunched: bool = false # true while training thread is running
|
|
roundCounter: int = 0
|
|
|
|
proc trainingThreadProc(args: TrainingArgs) {.thread.} =
|
|
var localAc = args.ac
|
|
var localAdam = args.adamStates
|
|
let m = ppoUpdate(localAc, args.buffer,
|
|
lastValue = args.lastValue,
|
|
adamStates = localAdam,
|
|
epochs = args.epochs,
|
|
miniBatchSize = args.miniBatchSize,
|
|
clipEpsilon = args.clipEpsilon,
|
|
entropyCoeff = args.entropyCoeff,
|
|
valueLossCoeff = args.valueLossCoeff,
|
|
lr = args.lr,
|
|
maxGradNorm = args.maxGradNorm,
|
|
gamma = args.gamma,
|
|
lam = args.lam)
|
|
saveCheckpoint(localAc, localAdam, args.weightsRoot, args.roundNum)
|
|
resultChan.send(TrainingResult(ac: localAc, adamStates: localAdam, metrics: m))
|
|
|
|
# ── 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()
|
|
bot.buffer = initTrajectoryBuffer()
|
|
bot.prevEnergy = 0.0'f32
|
|
bot.prevEnemyE = 0.0'f32
|
|
bot.hasLastTrans = false
|
|
bot.roundRewardSum = 0.0'f32
|
|
bot.roundTicks = 0
|
|
|
|
method onRoundEnded*(bot: PPOBot, e: RoundEndedEventForBot) =
|
|
inc roundCounter
|
|
|
|
# Add round-end score bonus to last transition (if any)
|
|
let roundReward = computeRoundReward(e.results.totalScore.float32)
|
|
if bot.hasLastTrans and bot.buffer.len > 0:
|
|
bot.buffer.transitions[^1].reward += roundReward
|
|
|
|
# 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 tr in bot.buffer.transitions: rewardSum += tr.reward
|
|
avgR = rewardSum / ticks.float32
|
|
let avgRStr = formatFloat(avgR.float, ffDecimal, 3)
|
|
printToStdOut(&"R:{roundCounter} ticks:{ticks} avgR:{avgRStr} score:{e.results.totalScore}\n")
|
|
echo &"R:{roundCounter} ticks:{ticks} avgR:{avgRStr} score:{e.results.totalScore}"
|
|
|
|
# Pick up result from previous training thread if available; channel IS the sync
|
|
if threadLaunched:
|
|
let (avail, trained) = resultChan.tryRecv()
|
|
if avail:
|
|
ac = trained.ac
|
|
gAdamStates = trained.adamStates
|
|
threadLaunched = false
|
|
let m = trained.metrics
|
|
printToStdOut(&" trained R:{roundCounter-1} 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-1} 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 ts = int(epochTime())
|
|
let jline = &"""{{\"type\":\"train\",\"round\":{roundCounter-1},\"actorLoss\":{m.actorLoss},\"valueLoss\":{m.valueLoss},\"gradNorm\":{m.gradNorm},\"ts\":{ts},{hp}}}"""
|
|
appendJsonLine(logFile, jline)
|
|
|
|
if bot.buffer.len == 0:
|
|
bot.hasLastTrans = false
|
|
return
|
|
|
|
# If last thread still running, drop this pass — start fresh with newer data
|
|
# ponytail: simple drop; queue if every round must train
|
|
if threadLaunched:
|
|
bot.buffer.clear()
|
|
bot.hasLastTrans = false
|
|
return
|
|
|
|
# Emit per-round game-stats JSON line (training health will follow when thread finishes)
|
|
let ts = int(epochTime())
|
|
let jline = &"""{{\"type\":\"round\",\"round\":{roundCounter},\"ticks\":{ticks},\"avgReward\":{avgR},\"score\":{e.results.totalScore},\"ts\":{ts}}}"""
|
|
appendJsonLine(logFile, jline)
|
|
|
|
let args = TrainingArgs(
|
|
ac: ac,
|
|
adamStates: gAdamStates,
|
|
buffer: bot.buffer,
|
|
lastValue: 0.0'f32,
|
|
roundNum: roundCounter,
|
|
weightsRoot: weightsRoot,
|
|
lr: hpLr,
|
|
clipEpsilon: hpClipEpsilon,
|
|
entropyCoeff: hpEntropyCoeff,
|
|
valueLossCoeff: hpValueLossCoeff,
|
|
maxGradNorm: hpMaxGradNorm,
|
|
gamma: hpGamma,
|
|
lam: hpLam,
|
|
epochs: hpEpochs,
|
|
miniBatchSize: hpMiniBatchSize,
|
|
)
|
|
bot.buffer.clear()
|
|
bot.hasLastTrans = false
|
|
|
|
printToStdOut(&" train→ R:{roundCounter} ticks:{ticks}\n")
|
|
echo &" train→ R:{roundCounter} ticks:{ticks}"
|
|
createThread(trainingThread, trainingThreadProc, args)
|
|
threadLaunched = true
|
|
|
|
method run(bot: PPOBot) =
|
|
# 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()
|
|
|
|
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)
|
|
let (rawActs, logP) = ac.actorForward(state)
|
|
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,
|
|
)
|
|
bot.buffer.add(tr)
|
|
|
|
# Store current for next tick
|
|
bot.lastState = state
|
|
bot.lastAction = 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:
|
|
resultChan.open()
|
|
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)
|