def testRolloutWithHeuristic(self, max_rollout_step, init_state,
gt_sumValue):
max_iteration = 1000
target_state = 6
isTerminal = Terminal(target_state)
catch_reward = 1
step_penalty = 0
reward_func = RewardFunction(step_penalty, catch_reward, isTerminal)
rolloutHeuristic = lambda state: 2
rollout_policy = lambda state: np.random.choice(self.action_space)
leaf_node = Node(id={1: init_state},
numVisited=1,
sumValue=0,
actionPrior=self.default_actionPrior,
isExpanded=True)
rollout = RollOut(rollout_policy, max_rollout_step, self.transition,
reward_func, isTerminal, rolloutHeuristic)
stored_reward = []
for curr_iter in range(max_iteration):
stored_reward.append(rollout(leaf_node))
calc_sumValue = np.mean(stored_reward)
self.assertAlmostEqual(gt_sumValue, calc_sumValue, places=1)
def mctsPolicy():
lowerBound = 0
gridSize = 10
upperBound = [gridSize - 1, gridSize - 1]
maxRunningSteps = 50
actionSpace = [(-1, 0), (1, 0), (0, 1), (0, -1)]
numActionSpace = len(actionSpace)
sheepSpeedRatio = 1
sheepActionSpace = list(map(tuple,
np.array(actionSpace) * sheepSpeedRatio))
numOfAgent = 5 # 1 hunter, 1 stag with high value, 3 rabbits with low value
hunterId = [0]
targetIds = [1, 2, 3, 4]
stagId = [1]
rabbitId = [2, 3, 4]
positionIndex = [0, 1]
getHunterPos = GetAgentPosFromState(hunterId, positionIndex)
getTargetsPos = GetAgentPosFromState(targetIds, positionIndex)
stayWithinBoundary = env.StayWithinBoundary(upperBound, lowerBound)
isTerminal = env.IsTerminal(getHunterPos, getTargetsPos)
transitionFunction = env.Transition(stayWithinBoundary)
reset = env.Reset(upperBound, lowerBound, numOfAgent)
# stagPolicy = RandomPolicy(sheepActionSpace)
stagPolicy = stationaryAgentPolicy
rabbitPolicies = [stationaryAgentPolicy] * len(rabbitId)
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
getActionPrior = lambda state: {
action: 1 / len(actionSpace)
for action in actionSpace
}
def wolfTransit(state, action):
return transitionFunction(
state, [action, maxFromDistribution(stagPolicy(state))] + [
maxFromDistribution(rabbitPolicy(state))
for rabbitPolicy in rabbitPolicies
])
stepPenalty = -1 / maxRunningSteps
catchBonus = 1
highRewardRatio = 10
rewardFunction = reward.RewardFunction(highRewardRatio, stepPenalty,
catchBonus, isTerminal)
initializeChildren = InitializeChildren(actionSpace, wolfTransit,
getActionPrior)
expand = Expand(isTerminal, initializeChildren)
def rolloutPolicy(state):
return actionSpace[np.random.choice(range(numActionSpace))]
rolloutHeuristicWeight = 0
rolloutHeuristic = reward.HeuristicDistanceToTarget(
rolloutHeuristicWeight, getHunterPos, getTargetsPos)
# rolloutHeuristic = lambda state: 0
maxRolloutSteps = 20
rollout = RollOut(rolloutPolicy, maxRolloutSteps, wolfTransit,
rewardFunction, isTerminal, rolloutHeuristic)
numSimulations = 600
wolfPolicy = MCTS(numSimulations, selectChild, expand, rollout, backup,
establishSoftmaxActionDist)
# All agents' policies
policy = lambda state: [
wolfPolicy(state), stagPolicy(state)
] + [rabbitPolicy(state) for rabbitPolicy in rabbitPolicies]
numOfAgent = 5
gridSize = 15
maxRunningSteps = 50
screenWidth = 600
screenHeight = 600
fullScreen = False
initializeScreen = InitializeScreen(screenWidth, screenHeight, fullScreen)
screen = initializeScreen()
# pg.mouse.set_visible(False)
leaveEdgeSpace = 2
lineWidth = 1
backgroundColor = [205, 255, 204]
lineColor = [0, 0, 0]
distractorColor = [255, 255, 0]
targetColor = [221, 160, 221]
playerColor = [50, 50, 255]
targetRadius = 10
playerRadius = 10
textColorTuple = (255, 50, 50)
drawBackground = DrawBackground(screen, gridSize, leaveEdgeSpace,
backgroundColor, lineColor, lineWidth,
textColorTuple)
drawNewState = DrawNewState(screen, drawBackground, targetColor,
playerColor, targetRadius, playerRadius)
chooseAction = [maxFromDistribution] * numOfAgent
renderOn = True
sampleTrajectory = SampleTrajectory(maxRunningSteps, transitionFunction,
isTerminal, reset, chooseAction,
renderOn, drawNewState)
startTime = time.time()
numOfEpisodes = 10
trajectories = [sampleTrajectory(policy) for i in range(numOfEpisodes)]
finshedTime = time.time() - startTime
print('lenght:', len(trajectories[0]))
print('time:', finshedTime)
def main():
DEBUG = 0
renderOn = 0
if DEBUG:
parametersForTrajectoryPath = {}
startSampleIndex = 5
endSampleIndex = 7
agentId = 1
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
else:
parametersForTrajectoryPath = json.loads(sys.argv[1])
startSampleIndex = int(sys.argv[2])
endSampleIndex = int(sys.argv[3])
agentId = int(parametersForTrajectoryPath['agentId'])
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
# check file exists or not
dirName = os.path.dirname(__file__)
trajectoriesSaveDirectory = os.path.join(
dirName, '..', '..', '..', '..', 'data', 'NoPhysics2wolves1sheep',
'trainWolvesTwoCenterControlAction88', 'trajectories')
if not os.path.exists(trajectoriesSaveDirectory):
os.makedirs(trajectoriesSaveDirectory)
trajectorySaveExtension = '.pickle'
maxRunningSteps = 50
numSimulations = 250
killzoneRadius = 150
fixedParameters = {
'agentId': agentId,
'maxRunningSteps': maxRunningSteps,
'numSimulations': numSimulations,
'killzoneRadius': killzoneRadius
}
generateTrajectorySavePath = GetSavePath(trajectoriesSaveDirectory,
trajectorySaveExtension,
fixedParameters)
trajectorySavePath = generateTrajectorySavePath(
parametersForTrajectoryPath)
if not os.path.isfile(trajectorySavePath):
numOfAgent = 3
xBoundary = [0, 600]
yBoundary = [0, 600]
resetState = Reset(xBoundary, yBoundary, numOfAgent)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(
xBoundary, yBoundary)
interpolateOneFrame = InterpolateOneFrame(
stayInBoundaryByReflectVelocity)
chooseInterpolatedNextState = lambda interpolatedStates: interpolatedStates[
-1]
sheepId = 0
wolvesId = 1
centerControlIndexList = [wolvesId]
unpackCenterControlAction = UnpackCenterControlAction(
centerControlIndexList)
numFramesToInterpolate = 0
transit = TransitWithInterpolation(numFramesToInterpolate,
interpolateOneFrame,
chooseInterpolatedNextState,
unpackCenterControlAction)
# NNGuidedMCTS init
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0), (-7, -7),
(0, -10), (7, -7), (0, 0)]
wolfActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0),
(-7, -7), (0, -10), (7, -7)]
preyPowerRatio = 10
sheepActionSpace = list(
map(tuple,
np.array(actionSpace) * preyPowerRatio))
predatorPowerRatio = 8
wolfActionOneSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolfActionTwoSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolvesActionSpace = list(
product(wolfActionOneSpace, wolfActionTwoSpace))
actionSpaceList = [sheepActionSpace, wolvesActionSpace]
# neural network init
numStateSpace = 2 * numOfAgent
numSheepActionSpace = len(sheepActionSpace)
numWolvesActionSpace = len(wolvesActionSpace)
regularizationFactor = 1e-4
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
generateSheepModel = GenerateModel(numStateSpace, numSheepActionSpace,
regularizationFactor)
# load save dir
NNModelSaveExtension = ''
sheepNNModelSaveDirectory = os.path.join(
dirName, '..', '..', '..', '..', 'data', 'NoPhysics2wolves1sheep',
'trainSheepWithTwoHeatSeekingWolves', 'trainedResNNModels')
sheepNNModelFixedParameters = {
'agentId': 0,
'maxRunningSteps': 50,
'numSimulations': 110,
'miniBatchSize': 256,
'learningRate': 0.0001,
}
getSheepNNModelSavePath = GetSavePath(sheepNNModelSaveDirectory,
NNModelSaveExtension,
sheepNNModelFixedParameters)
depth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepNNModel = generateSheepModel(sharedWidths * depth,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod,
dropoutRate)
sheepTrainedModelPath = getSheepNNModelSavePath({
'trainSteps': 50000,
'depth': depth
})
sheepTrainedModel = restoreVariables(initSheepNNModel,
sheepTrainedModelPath)
sheepPolicy = ApproximatePolicy(sheepTrainedModel, sheepActionSpace)
wolfOneId = 1
wolfTwoId = 2
xPosIndex = [0, 1]
getSheepXPos = GetAgentPosFromState(sheepId, xPosIndex)
getWolfOneXPos = GetAgentPosFromState(wolfOneId, xPosIndex)
speed = 120
#sheepPolicy = HeatSeekingContinuesDeterministicPolicy(getWolfOneXPos, getSheepXPos, speed)
# MCTS
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
# prior
getActionPrior = lambda state: {
action: 1 / len(wolvesActionSpace)
for action in wolvesActionSpace
}
# load chase nn policy
chooseActionInMCTS = sampleFromDistribution
def wolvesTransit(state, action):
return transit(state,
[chooseActionInMCTS(sheepPolicy(state)), action])
# reward function
wolfOneId = 1
wolfTwoId = 2
xPosIndex = [0, 1]
getSheepXPos = GetAgentPosFromState(sheepId, xPosIndex)
getWolfOneXPos = GetAgentPosFromState(wolfOneId, xPosIndex)
getWolfTwoXPos = GetAgentPosFromState(wolfTwoId, xPosIndex)
isCollidedOne = IsTerminal(getWolfOneXPos, getSheepXPos,
killzoneRadius)
isCollidedTwo = IsTerminal(getWolfTwoXPos, getSheepXPos,
killzoneRadius)
calCollisionTimes = lambda state: np.sum([
isCollidedOne(state), isCollidedTwo(state)
]) # collisionTimeByAddingCollisionInAllWolves
#calCollisionTimes = lambda state: np.max([isCollidedOne(state), isCollidedTwo(state)]) # collisionTimeByBooleanCollisionForAnyWolf
calTerminationSignals = calCollisionTimes
chooseInterpolatedStateByEarlyTermination = ChooseInterpolatedStateByEarlyTermination(
calTerminationSignals)
numFramesToInterpolateInReward = 3
interpolateStateInReward = TransitWithInterpolation(
numFramesToInterpolateInReward, interpolateOneFrame,
chooseInterpolatedStateByEarlyTermination,
unpackCenterControlAction)
aliveBonus = -1 / maxRunningSteps * 10
deathPenalty = 1
rewardFunction = RewardFunctionCompeteWithStateInterpolation(
aliveBonus, deathPenalty, calCollisionTimes,
interpolateStateInReward)
# initialize children; expand
initializeChildren = InitializeChildren(wolvesActionSpace,
wolvesTransit, getActionPrior)
isTerminal = lambda state: False
expand = Expand(isTerminal, initializeChildren)
# random rollout policy
def rolloutPolicy(state):
return [
sampleFromDistribution(sheepPolicy(state)),
wolvesActionSpace[np.random.choice(
range(numWolvesActionSpace))]
]
# rollout
#rolloutHeuristicWeight = 0
#minDistance = 400
#rolloutHeuristic1 = HeuristicDistanceToTarget(
# rolloutHeuristicWeight, getWolfOneXPos, getSheepXPos, minDistance)
#rolloutHeuristic2 = HeuristicDistanceToTarget(
# rolloutHeuristicWeight, getWolfTwoXPos, getSheepXPos, minDistance)
#rolloutHeuristic = lambda state: (rolloutHeuristic1(state) + rolloutHeuristic2(state)) / 2
rolloutHeuristic = lambda state: 0
maxRolloutSteps = 15
rollout = RollOut(rolloutPolicy, maxRolloutSteps, transit,
rewardFunction, isTerminal, rolloutHeuristic)
wolfPolicy = MCTS(numSimulations, selectChild, expand, rollout, backup,
establishSoftmaxActionDist)
# All agents' policies
policy = lambda state: [sheepPolicy(state), wolfPolicy(state)]
chooseActionList = [maxFromDistribution, maxFromDistribution]
def sampleAction(state):
actionDists = [sheepPolicy(state), wolfPolicy(state)]
action = [
chooseAction(actionDist) for actionDist, chooseAction in zip(
actionDists, chooseActionList)
]
return action
render = None
if renderOn:
import pygame as pg
from pygame.color import THECOLORS
screenColor = THECOLORS['black']
circleColorList = [
THECOLORS['green'], THECOLORS['yellow'], THECOLORS['red']
]
circleSize = 10
saveImage = False
saveImageDir = os.path.join(dirName, '..', '..', '..', '..',
'data', 'demoImg')
if not os.path.exists(saveImageDir):
os.makedirs(saveImageDir)
screen = pg.display.set_mode([max(xBoundary), max(yBoundary)])
render = Render(numOfAgent, xPosIndex, screen, screenColor,
circleColorList, circleSize, saveImage,
saveImageDir)
forwardOneStep = ForwardOneStep(transit, rewardFunction)
sampleTrajectory = SampleTrajectoryWithRender(maxRunningSteps,
isTerminal, resetState,
forwardOneStep, render,
renderOn)
trajectories = [
sampleTrajectory(sampleAction)
for sampleIndex in range(startSampleIndex, endSampleIndex)
]
print([len(traj) for traj in trajectories])
saveToPickle(trajectories, trajectorySavePath)
def main():
startTime = time.time()
DEBUG = 1
renderOn = 1
if DEBUG:
parametersForTrajectoryPath = {}
startSampleIndex = 5
endSampleIndex = 8
agentId = 1
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
else:
parametersForTrajectoryPath = json.loads(sys.argv[1])
startSampleIndex = int(sys.argv[2])
endSampleIndex = int(sys.argv[3])
agentId = int(parametersForTrajectoryPath['agentId'])
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
# check file exists or not
dirName = os.path.dirname(__file__)
trajectoriesSaveDirectory = os.path.join(
dirName, '..', '..', '..', '..', 'data', 'MADDPG2wolves1sheep',
'trainWolvesTwoCenterControlAction', 'trajectories')
if not os.path.exists(trajectoriesSaveDirectory):
os.makedirs(trajectoriesSaveDirectory)
trajectorySaveExtension = '.pickle'
maxRunningSteps = 50
numSimulations = 250
fixedParameters = {
'agentId': agentId,
'maxRunningSteps': maxRunningSteps,
'numSimulations': numSimulations
}
generateTrajectorySavePath = GetSavePath(trajectoriesSaveDirectory,
trajectorySaveExtension,
fixedParameters)
trajectorySavePath = generateTrajectorySavePath(
parametersForTrajectoryPath)
if not os.path.isfile(trajectorySavePath):
# env MDP
sheepsID = [0]
wolvesID = [1, 2]
blocksID = []
numSheeps = len(sheepsID)
numWolves = len(wolvesID)
numBlocks = len(blocksID)
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
sheepSize = 0.05
wolfSize = 0.075
blockSize = 0.2
sheepMaxSpeed = 1.3 * 1
wolfMaxSpeed = 1.0 * 1
blockMaxSpeed = None
entitiesSizeList = [sheepSize] * numSheeps + [wolfSize] * numWolves + [
blockSize
] * numBlocks
entityMaxSpeedList = [sheepMaxSpeed] * numSheeps + [
wolfMaxSpeed
] * numWolves + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
centralControlId = 1
centerControlIndexList = [centralControlId]
reshapeAction = UnpackCenterControlAction(centerControlIndexList)
getCollisionForce = GetCollisionForce()
applyActionForce = ApplyActionForce(wolvesID, sheepsID,
entitiesMovableList)
applyEnvironForce = ApplyEnvironForce(numEntities, entitiesMovableList,
entitiesSizeList,
getCollisionForce,
getPosFromAgentState)
integrateState = IntegrateState(numEntities, entitiesMovableList,
massList, entityMaxSpeedList,
getVelFromAgentState,
getPosFromAgentState)
interpolateState = TransitMultiAgentChasing(numEntities, reshapeAction,
applyActionForce,
applyEnvironForce,
integrateState)
numFramesToInterpolate = 1
def transit(state, action):
for frameIndex in range(numFramesToInterpolate):
nextState = interpolateState(state, action)
action = np.array([(0, 0)] * numAgents)
state = nextState
return nextState
isTerminal = lambda state: False
isCollision = IsCollision(getPosFromAgentState)
collisonRewardWolf = 1
punishForOutOfBound = PunishForOutOfBound()
rewardWolf = RewardCentralControlPunishBond(
wolvesID, sheepsID, entitiesSizeList, getPosFromAgentState,
isCollision, punishForOutOfBound, collisonRewardWolf)
collisonRewardSheep = -1
rewardSheep = RewardCentralControlPunishBond(
sheepsID, wolvesID, entitiesSizeList, getPosFromAgentState,
isCollision, punishForOutOfBound, collisonRewardSheep)
resetState = ResetMultiAgentChasing(numAgents, numBlocks)
observeOneAgent = lambda agentID: Observe(
agentID, wolvesID, sheepsID, blocksID, getPosFromAgentState,
getVelFromAgentState)
observe = lambda state: [
observeOneAgent(agentID)(state) for agentID in range(numAgents)
]
# policy
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0), (-7, -7),
(0, -10), (7, -7), (0, 0)]
wolfActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0),
(-7, -7), (0, -10), (7, -7), (0, 0)]
preyPowerRatio = 0.5
sheepActionSpace = list(
map(tuple,
np.array(actionSpace) * preyPowerRatio))
predatorPowerRatio = 0.5
wolfActionOneSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolfActionTwoSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolvesActionSpace = list(
product(wolfActionOneSpace, wolfActionTwoSpace))
actionSpaceList = [sheepActionSpace, wolvesActionSpace]
# neural network init
numStateSpace = 4 * numEntities
numSheepActionSpace = len(sheepActionSpace)
numWolvesActionSpace = len(wolvesActionSpace)
regularizationFactor = 1e-4
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
generateSheepModel = GenerateModel(numStateSpace, numSheepActionSpace,
regularizationFactor)
sheepPolicy = lambda state: {(0, 0): 1}
# MCTS
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
# prior
getActionPrior = lambda state: {
action: 1 / len(wolvesActionSpace)
for action in wolvesActionSpace
}
# load chase nn policy
chooseActionInMCTS = sampleFromDistribution
def wolvesTransit(state, action):
return transit(state,
[chooseActionInMCTS(sheepPolicy(state)), action])
# initialize children; expand
initializeChildren = InitializeChildren(wolvesActionSpace,
wolvesTransit, getActionPrior)
isTerminal = lambda state: False
expand = Expand(isTerminal, initializeChildren)
# random rollout policy
def rolloutPolicy(state):
return [
sampleFromDistribution(sheepPolicy(state)),
wolvesActionSpace[np.random.choice(
range(numWolvesActionSpace))]
]
rolloutHeuristic = lambda state: 0
maxRolloutSteps = 15
rollout = RollOut(rolloutPolicy, maxRolloutSteps, transit, rewardWolf,
isTerminal, rolloutHeuristic)
wolfPolicy = MCTS(numSimulations, selectChild, expand, rollout, backup,
establishSoftmaxActionDist)
# All agents' policies
policy = lambda state: [sheepPolicy(state), wolfPolicy(state)]
chooseActionList = [maxFromDistribution, maxFromDistribution]
def sampleAction(state):
actionDists = [sheepPolicy(state), wolfPolicy(state)]
action = [
chooseAction(actionDist) for actionDist, chooseAction in zip(
actionDists, chooseActionList)
]
return action
render = None
forwardOneStep = ForwardOneStep(transit, rewardWolf)
sampleTrajectory = SampleTrajectoryWithRender(maxRunningSteps,
isTerminal, resetState,
forwardOneStep, render,
renderOn)
trajectories = [
sampleTrajectory(sampleAction)
for sampleIndex in range(startSampleIndex, endSampleIndex)
]
print([len(traj) for traj in trajectories])
saveToPickle(trajectories, trajectorySavePath)
endTime = time.time()
def main():
DEBUG = 0
renderOn = 0
if DEBUG:
parametersForTrajectoryPath = {}
startSampleIndex = 0
endSampleIndex = 10
agentId = 1
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
else:
parametersForTrajectoryPath = json.loads(sys.argv[1])
startSampleIndex = int(sys.argv[2])
endSampleIndex = int(sys.argv[3])
agentId = int(parametersForTrajectoryPath['agentId'])
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex,
endSampleIndex)
# check file exists or not
dirName = os.path.dirname(__file__)
trajectoriesSaveDirectory = os.path.join(
dirName, '..', '..', '..', '..', 'data', '2wolves1sheep',
'trainWolvesTwoCenterControlMultiTrees', 'trajectories')
if not os.path.exists(trajectoriesSaveDirectory):
os.makedirs(trajectoriesSaveDirectory)
trajectorySaveExtension = '.pickle'
maxRunningSteps = 50
numSimulations = 500
killzoneRadius = 50
fixedParameters = {
'agentId': agentId,
'maxRunningSteps': maxRunningSteps,
'numSimulations': numSimulations,
'killzoneRadius': killzoneRadius
}
generateTrajectorySavePath = GetSavePath(trajectoriesSaveDirectory,
trajectorySaveExtension,
fixedParameters)
trajectorySavePath = generateTrajectorySavePath(
parametersForTrajectoryPath)
if not os.path.isfile(trajectorySavePath):
numOfAgent = 3
sheepId = 0
wolvesId = 1
wolfOneId = 1
wolfTwoId = 2
xPosIndex = [0, 1]
xBoundary = [0, 600]
yBoundary = [0, 600]
getSheepXPos = GetAgentPosFromState(sheepId, xPosIndex)
getWolfOneXPos = GetAgentPosFromState(wolfOneId, xPosIndex)
getWolfTwoXPos = GetAgentPosFromState(wolfTwoId, xPosIndex)
reset = Reset(xBoundary, yBoundary, numOfAgent)
isTerminalOne = IsTerminal(getWolfOneXPos, getSheepXPos,
killzoneRadius)
isTerminalTwo = IsTerminal(getWolfTwoXPos, getSheepXPos,
killzoneRadius)
isTerminal = lambda state: isTerminalOne(state) or isTerminalTwo(state)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(
xBoundary, yBoundary)
centerControlIndexList = [wolvesId]
unpackCenterControlAction = UnpackCenterControlAction(
centerControlIndexList)
transitionFunction = TransiteForNoPhysicsWithCenterControlAction(
stayInBoundaryByReflectVelocity)
numFramesToInterpolate = 3
transit = TransitWithInterpolateStateWithCenterControlAction(
numFramesToInterpolate, transitionFunction, isTerminal,
unpackCenterControlAction)
# NNGuidedMCTS init
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0), (-7, -7),
(0, -10), (7, -7), (0, 0)]
wolfActionSpace = actionSpace
# wolfActionSpace = [(10, 0), (0, 10), (-10, 0), (0, -10), (0, 0)]
preyPowerRatio = 12
sheepActionSpace = list(
map(tuple,
np.array(actionSpace) * preyPowerRatio))
predatorPowerRatio = 8
wolfActionOneSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolfActionTwoSpace = list(
map(tuple,
np.array(wolfActionSpace) * predatorPowerRatio))
wolvesActionSpace = list(
product(wolfActionOneSpace, wolfActionTwoSpace))
actionSpaceList = [sheepActionSpace, wolvesActionSpace]
# neural network init
numStateSpace = 2 * numOfAgent
numSheepActionSpace = len(sheepActionSpace)
numWolvesActionSpace = len(wolvesActionSpace)
regularizationFactor = 1e-4
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
generateSheepModel = GenerateModel(numStateSpace, numSheepActionSpace,
regularizationFactor)
# load save dir
NNModelSaveExtension = ''
sheepNNModelSaveDirectory = os.path.join(
dirName, '..', '..', '..', '..', 'data', '2wolves1sheep',
'trainSheepWithTwoHeatSeekingWolves', 'trainedResNNModels')
sheepNNModelFixedParameters = {
'agentId': 0,
'maxRunningSteps': 50,
'numSimulations': 110,
'miniBatchSize': 256,
'learningRate': 0.0001,
}
getSheepNNModelSavePath = GetSavePath(sheepNNModelSaveDirectory,
NNModelSaveExtension,
sheepNNModelFixedParameters)
depth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepNNModel = generateSheepModel(sharedWidths * depth,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod,
dropoutRate)
sheepTrainedModelPath = getSheepNNModelSavePath({
'trainSteps': 50000,
'depth': depth
})
sheepTrainedModel = restoreVariables(initSheepNNModel,
sheepTrainedModelPath)
sheepPolicy = ApproximatePolicy(sheepTrainedModel, sheepActionSpace)
# MCTS
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
# prior
getActionPrior = lambda state: {
action: 1 / len(wolvesActionSpace)
for action in wolvesActionSpace
}
# load chase nn policy
temperatureInMCTS = 1
chooseActionInMCTS = SampleAction(temperatureInMCTS)
def wolvesTransit(state, action):
return transit(state,
[chooseActionInMCTS(sheepPolicy(state)), action])
# reward function
aliveBonus = -1 / maxRunningSteps
deathPenalty = 1
rewardFunction = reward.RewardFunctionCompete(aliveBonus, deathPenalty,
isTerminal)
# initialize children; expand
initializeChildren = InitializeChildren(wolvesActionSpace,
wolvesTransit, getActionPrior)
expand = Expand(isTerminal, initializeChildren)
# random rollout policy
def rolloutPolicy(state):
return wolvesActionSpace[np.random.choice(
range(numWolvesActionSpace))]
# rollout
rolloutHeuristicWeight = 0
minDistance = 400
rolloutHeuristic1 = reward.HeuristicDistanceToTarget(
rolloutHeuristicWeight, getWolfOneXPos, getSheepXPos, minDistance)
rolloutHeuristic2 = reward.HeuristicDistanceToTarget(
rolloutHeuristicWeight, getWolfTwoXPos, getSheepXPos, minDistance)
rolloutHeuristic = lambda state: (rolloutHeuristic1(state) +
rolloutHeuristic2(state)) / 2
maxRolloutSteps = 15
rollout = RollOut(rolloutPolicy, maxRolloutSteps, wolvesTransit,
rewardFunction, isTerminal, rolloutHeuristic)
numTree = 4
numSimulationsPerTree = int(numSimulations / numTree)
wolfPolicy = StochasticMCTS(
numTree, numSimulationsPerTree, selectChild, expand, rollout,
backup, establishSoftmaxActionDistFromMultipleTrees)
# All agents' policies
policy = lambda state: [sheepPolicy(state), wolfPolicy(state)]
chooseActionList = [chooseGreedyAction, chooseGreedyAction]
render = None
if renderOn:
import pygame as pg
from pygame.color import THECOLORS
screenColor = THECOLORS['black']
circleColorList = [
THECOLORS['green'], THECOLORS['red'], THECOLORS['red']
]
circleSize = 10
saveImage = False
saveImageDir = os.path.join(dirName, '..', '..', '..', '..',
'data', 'demoImg')
if not os.path.exists(saveImageDir):
os.makedirs(saveImageDir)
screen = pg.display.set_mode([xBoundary[1], yBoundary[1]])
render = Render(numOfAgent, xPosIndex, screen, screenColor,
circleColorList, circleSize, saveImage,
saveImageDir)
sampleTrajectory = SampleTrajectoryWithRender(maxRunningSteps, transit,
isTerminal, reset,
chooseActionList, render,
renderOn)
trajectories = [
sampleTrajectory(policy)
for sampleIndex in range(startSampleIndex, endSampleIndex)
]
print([len(traj) for traj in trajectories])
saveToPickle(trajectories, trajectorySavePath)