def __call__(self, parameters):
print(parameters)
numWolves = parameters['numWolves']
numSheep = 1
## MDP Env
# state is all multi agent state # action is all multi agent action
xBoundary = [0,600]
yBoundary = [0,600]
numOfAgent = numWolves + numSheep
reset = Reset(xBoundary, yBoundary, numOfAgent)
possibleSheepIds = list(range(numSheep))
possibleWolvesIds = list(range(numSheep, numSheep + numWolves))
getSheepStatesFromAll = lambda state: np.array(state)[possibleSheepIds]
getWolvesStatesFromAll = lambda state: np.array(state)[possibleWolvesIds]
killzoneRadius = 50
isTerminal = IsTerminal(killzoneRadius, getSheepStatesFromAll, getWolvesStatesFromAll)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(xBoundary, yBoundary)
interpolateOneFrame = InterpolateOneFrame(stayInBoundaryByReflectVelocity)
numFramesToInterpolate = 3
transit = TransitWithTerminalCheckOfInterpolation(numFramesToInterpolate, interpolateOneFrame, isTerminal)
maxRunningSteps = 52
timeCost = 1/maxRunningSteps
terminalBonus = 1
rewardFunction = RewardFunctionByTerminal(timeCost, terminalBonus, isTerminal)
forwardOneStep = ForwardOneStep(transit, rewardFunction)
sampleTrajectory = SampleTrajectory(maxRunningSteps, isTerminal, reset, forwardOneStep)
## MDP Policy
# Sheep Part
# Sheep Policy Function
numSheepPolicyStateSpace = 2 * (numWolves + 1)
sheepActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7),
(-10, 0), (-7, -7), (0, -10), (7, -7), (0, 0)]
preyPowerRatio = 12
sheepIndividualActionSpace = list(map(tuple, np.array(sheepActionSpace) * preyPowerRatio))
numSheepActionSpace = len(sheepIndividualActionSpace)
regularizationFactor = 1e-4
generateSheepModel = GenerateModel(numSheepPolicyStateSpace, numSheepActionSpace, regularizationFactor)
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
sheepNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepModel = generateSheepModel(sharedWidths * sheepNNDepth, actionLayerWidths, valueLayerWidths,
resBlockSize, initializationMethod, dropoutRate)
sheepModelPath = os.path.join('..', '..', 'data', 'preTrainModel',
'agentId=0.'+str(numWolves)+'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations=110_trainSteps=50000')
sheepNNModel = restoreVariables(initSheepModel, sheepModelPath)
sheepPolicy = ApproximatePolicy(sheepNNModel, sheepIndividualActionSpace)
# Sheep Generate Action
softParameterInPlanningForSheep = 2.5
softPolicyInPlanningForSheep = SoftDistribution(softParameterInPlanningForSheep)
softenSheepPolicy = lambda relativeAgentsStatesForSheepPolicy: softPolicyInPlanningForSheep(sheepPolicy(relativeAgentsStatesForSheepPolicy))
sheepChooseActionMethod = sampleFromDistribution
sheepSampleActions = [SampleActionOnFixedIntention(selfId, possibleWolvesIds, sheepPolicy, sheepChooseActionMethod) for selfId in possibleSheepIds]
# Wolves Part
# Policy Likelihood function: Wolf Centrol Control NN Policy Given Intention
numWolvesStateSpaces = [2 * (numInWe + 1)
for numInWe in range(2, numWolves + 1)]
actionSpace = [(10, 0), (0, 10), (-10, 0), (0, -10)]
predatorPowerRatio = 8
wolfIndividualActionSpace = list(map(tuple, np.array(actionSpace) * predatorPowerRatio))
wolvesCentralControlActionSpaces = [list(it.product(wolfIndividualActionSpace, repeat = numInWe))
for numInWe in range(2, numWolves + 1)]
numWolvesCentralControlActionSpaces = [len(wolvesCentralControlActionSpace)
for wolvesCentralControlActionSpace in wolvesCentralControlActionSpaces]
regularizationFactor = 1e-4
generateWolvesCentralControlModels = [GenerateModel(numStateSpace, numActionSpace, regularizationFactor)
for numStateSpace, numActionSpace in zip(numWolvesStateSpaces, numWolvesCentralControlActionSpaces)]
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
wolfNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initWolvesCentralControlModels = [generateWolvesCentralControlModel(sharedWidths * wolfNNDepth, actionLayerWidths, valueLayerWidths,
resBlockSize, initializationMethod, dropoutRate) for generateWolvesCentralControlModel in generateWolvesCentralControlModels]
NNNumSimulations = 250
wolvesModelPaths = [os.path.join('..', '..', 'data', 'preTrainModel',
'agentId='+str(len(actionSpace) * np.sum([10**_ for _ in
range(numInWe)]))+'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations='+str(NNNumSimulations)+'_trainSteps=50000')
for numInWe in range(2, numWolves + 1)]
print(wolvesModelPaths)
wolvesCentralControlNNModels = [restoreVariables(initWolvesCentralControlModel, wolvesModelPath)
for initWolvesCentralControlModel, wolvesModelPath in zip(initWolvesCentralControlModels, wolvesModelPaths)]
wolvesCentralControlPolicies = [ApproximatePolicy(NNModel, actionSpace)
for NNModel, actionSpace in zip(wolvesCentralControlNNModels, wolvesCentralControlActionSpaces)]
centralControlPolicyListBasedOnNumAgentsInWe = wolvesCentralControlPolicies # 0 for two agents in We, 1 for three agents...
softParameterInInference = 1
softPolicyInInference = SoftDistribution(softParameterInInference)
policyForCommittedAgentsInInference = PolicyForCommittedAgent(centralControlPolicyListBasedOnNumAgentsInWe, softPolicyInInference,
getStateThirdPersonPerspective)
calCommittedAgentsPolicyLikelihood = CalCommittedAgentsPolicyLikelihood(policyForCommittedAgentsInInference)
wolfLevel2ActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7),
(-10, 0), (-7, -7), (0, -10), (7, -7)]
wolfLevel2IndividualActionSpace = list(map(tuple, np.array(wolfLevel2ActionSpace) * predatorPowerRatio))
wolfLevel2CentralControlActionSpace = list(it.product(wolfLevel2IndividualActionSpace))
numWolfLevel2ActionSpace = len(wolfLevel2CentralControlActionSpace)
regularizationFactor = 1e-4
generatewolfLevel2Models = [GenerateModel(numStateSpace, numWolfLevel2ActionSpace, regularizationFactor) for numStateSpace in numWolvesStateSpaces]
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
wolfLevel2NNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initwolfLevel2Models = [generatewolfLevel2Model(sharedWidths * wolfLevel2NNDepth, actionLayerWidths, valueLayerWidths,
resBlockSize, initializationMethod, dropoutRate) for generatewolfLevel2Model in generatewolfLevel2Models]
wolfLevel2ModelPaths = [os.path.join('..', '..', 'data', 'preTrainModel',
'agentId=1.'+str(numInWe)+'_depth=9_hierarchy=2_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations='+str(NNNumSimulations)+'_trainSteps=50000')
for numInWe in range(2, numWolves + 1)]
wolfLevel2NNModels = [restoreVariables(initwolfLevel2Model, wolfLevel2ModelPath)
for initwolfLevel2Model, wolfLevel2ModelPath in zip(initwolfLevel2Models, wolfLevel2ModelPaths)]
wolfLevel2Policies = [ApproximatePolicy(wolfLevel2NNModel, wolfLevel2CentralControlActionSpace)
for wolfLevel2NNModel in wolfLevel2NNModels]
level2PolicyListBasedOnNumAgentsInWe = wolfLevel2Policies # 0 for two agents in We, 1 for three agents...
softPolicy = SoftDistribution(2.5)
totalInSmallRangeFlags = []
for trial in range(self.numTrajectories):
state = reset()
while isTerminal(state):
state = reset()
jointActions = sampleFromDistribution(softPolicy(wolvesCentralControlPolicies[numWolves - 2](state)))
hierarchyActions = []
weIds = [list(range(numSheep, numWolves + numSheep)) for _ in range(numWolves)]
for index in range(numWolves):
weId = weIds[index].copy()
weId.insert(0, weId.pop(index))
relativeId = [0] + weId
action = sampleFromDistribution(softPolicy(wolfLevel2Policies[numWolves - 2](state[relativeId])))
hierarchyActions.append(action)
reasonableActionRange = [int(np.linalg.norm(np.array(jointAction) - np.array(hierarchyAction)) <= 8 * predatorPowerRatio)
for jointAction, hierarchyAction in zip(jointActions, hierarchyActions) if jointAction != (0, 0) and hierarchyAction != (0, 0)]
totalInSmallRangeFlags = totalInSmallRangeFlags + reasonableActionRange
inSmallRangeRateMean = np.mean(totalInSmallRangeFlags)
return inSmallRangeRateMean
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 __call__(self, parameters):
print(parameters)
numWolves = parameters['numWolves']
numSheep = parameters['numSheep']
softParameterInInference = parameters['inferenceSoft']
softParameterInPlanning = parameters['wolfPolicySoft']
otherCompeteRate = parameters['otherCompeteRate']
competeDetectionRate = parameters['competeDetectionRate']
## MDP Env
# state is all multi agent state # action is all multi agent action
xBoundary = [0, 600]
yBoundary = [0, 600]
numOfAgent = numWolves + numSheep
reset = Reset(xBoundary, yBoundary, numOfAgent)
possibleSheepIds = list(range(numSheep))
possibleWolvesIds = list(range(numSheep, numSheep + numWolves))
getSheepStatesFromAll = lambda state: np.array(state)[possibleSheepIds]
getWolvesStatesFromAll = lambda state: np.array(state)[
possibleWolvesIds]
killzoneRadius = 50
isTerminal = IsTerminal(killzoneRadius, getSheepStatesFromAll,
getWolvesStatesFromAll)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(
xBoundary, yBoundary)
interpolateOneFrame = InterpolateOneFrame(
stayInBoundaryByReflectVelocity)
numFramesToInterpolate = 3
transit = TransitWithTerminalCheckOfInterpolation(
numFramesToInterpolate, interpolateOneFrame, isTerminal)
maxRunningSteps = 61
timeCost = 1 / maxRunningSteps
terminalBonus = 1
rewardFunction = RewardFunctionByTerminal(timeCost, terminalBonus,
isTerminal)
forwardOneStep = ForwardOneStep(transit, rewardFunction)
sampleTrajectory = SampleTrajectory(maxRunningSteps, isTerminal, reset,
forwardOneStep)
## MDP Policy
# Sheep Part
# Sheep Policy Function
numSheepPolicyStateSpace = 2 * (numWolves + 1)
sheepActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0),
(-7, -7), (0, -10), (7, -7), (0, 0)]
preyPowerRatio = 12
sheepIndividualActionSpace = list(
map(tuple,
np.array(sheepActionSpace) * preyPowerRatio))
numSheepActionSpace = len(sheepIndividualActionSpace)
regularizationFactor = 1e-4
generateSheepModel = GenerateModel(numSheepPolicyStateSpace,
numSheepActionSpace,
regularizationFactor)
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
sheepNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepModel = generateSheepModel(sharedWidths * sheepNNDepth,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod, dropoutRate)
sheepModelPath = os.path.join(
'..', '..', 'data', 'preTrainModel',
'agentId=0.' + str(numWolves) +
'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations=110_trainSteps=50000'
)
sheepNNModel = restoreVariables(initSheepModel, sheepModelPath)
sheepPolicy = ApproximatePolicy(sheepNNModel,
sheepIndividualActionSpace)
# Sheep Generate Action
softParameterInPlanningForSheep = 2.0
softPolicyInPlanningForSheep = SoftDistribution(
softParameterInPlanningForSheep)
softenSheepPolicy = lambda relativeAgentsStatesForSheepPolicy: softPolicyInPlanningForSheep(
sheepPolicy(relativeAgentsStatesForSheepPolicy))
sheepChooseActionMethod = sampleFromDistribution
sheepSampleActions = [
SampleActionOnFixedIntention(selfId, possibleWolvesIds,
softenSheepPolicy,
sheepChooseActionMethod)
for selfId in possibleSheepIds
]
# Wolves Part
# Percept Action For Inference
perceptAction = lambda action: action
# Policy Likelihood function: Wolf Centrol Control NN Policy Given Intention
numWolvesStateSpaces = [
2 * (numInWe + 1) for numInWe in range(2, numWolves + 1)
]
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0), (-7, -7),
(0, -10), (7, -7)]
predatorPowerRatio = 8
wolfIndividualActionSpace = list(
map(tuple,
np.array(actionSpace) * predatorPowerRatio))
wolvesCentralControlActionSpaces = [
list(it.product(wolfIndividualActionSpace, repeat=numInWe))
for numInWe in range(2, numWolves + 1)
]
numWolvesCentralControlActionSpaces = [
len(wolvesCentralControlActionSpace) for
wolvesCentralControlActionSpace in wolvesCentralControlActionSpaces
]
regularizationFactor = 1e-4
generateWolvesCentralControlModels = [
GenerateModel(numStateSpace, numActionSpace, regularizationFactor)
for numStateSpace, numActionSpace in zip(
numWolvesStateSpaces, numWolvesCentralControlActionSpaces)
]
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
wolfNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initWolvesCentralControlModels = [
generateWolvesCentralControlModel(sharedWidths * wolfNNDepth,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod,
dropoutRate)
for generateWolvesCentralControlModel in
generateWolvesCentralControlModels
]
NNNumSimulations = 250
wolvesModelPaths = [
os.path.join(
'..', '..', 'data', 'preTrainModel',
'agentId=' + str(8 * np.sum([10**_ for _ in range(numInWe)])) +
'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations='
+ str(NNNumSimulations) + '_trainSteps=50000')
for numInWe in range(2, numWolves + 1)
]
print(wolvesModelPaths)
wolvesCentralControlNNModels = [
restoreVariables(initWolvesCentralControlModel, wolvesModelPath)
for initWolvesCentralControlModel, wolvesModelPath in zip(
initWolvesCentralControlModels, wolvesModelPaths)
]
wolvesCentralControlPolicies = [
ApproximatePolicy(NNModel, actionSpace) for NNModel, actionSpace in
zip(wolvesCentralControlNNModels, wolvesCentralControlActionSpaces)
]
centralControlPolicyListBasedOnNumAgentsInWe = wolvesCentralControlPolicies # 0 for two agents in We, 1 for three agents...
softPolicyInInference = SoftDistribution(softParameterInInference)
policyForCommittedAgentsInInference = PolicyForCommittedAgent(
centralControlPolicyListBasedOnNumAgentsInWe,
softPolicyInInference, getStateOrActionThirdPersonPerspective)
concernedAgentsIds = [2]
calCommittedAgentsPolicyLikelihood = CalCommittedAgentsPolicyLikelihood(
concernedAgentsIds, policyForCommittedAgentsInInference)
getGoalStateForIndividualHeatseeking = lambda statesRelative: np.array(
statesRelative)[0]
getSelfStateForIndividualHeatseeking = lambda statesRelative: np.array(
statesRelative)[1]
heatseekingPrecesion = 1.83
heatSeekingDiscreteStochasticPolicy = HeatSeekingDiscreteStochasticPolicy(
heatseekingPrecesion, wolfIndividualActionSpace,
getSelfStateForIndividualHeatseeking,
getGoalStateForIndividualHeatseeking)
policyForUncommittedAgentsInInference = PolicyForUncommittedAgent(
possibleWolvesIds, heatSeekingDiscreteStochasticPolicy,
softPolicyInInference, getStateOrActionFirstPersonPerspective)
calUncommittedAgentsPolicyLikelihood = CalUncommittedAgentsPolicyLikelihood(
possibleWolvesIds, concernedAgentsIds,
policyForUncommittedAgentsInInference)
# Joint Likelihood
calJointLikelihood = lambda intention, state, perceivedAction: calCommittedAgentsPolicyLikelihood(intention, state, perceivedAction) * \
calUncommittedAgentsPolicyLikelihood(intention, state, perceivedAction)
wolvesValueListBasedOnNumAgentsInWe = [
ApproximateValue(NNModel)
for NNModel in wolvesCentralControlNNModels
]
calIntentionValueGivenState = CalIntentionValueGivenState(
wolvesValueListBasedOnNumAgentsInWe)
softParamterForValue = 0.01
softValueToBuildDistribution = SoftMax(softParamterForValue)
adjustIntentionPriorGivenValueOfState = AdjustIntentionPriorGivenValueOfState(
calIntentionValueGivenState, softValueToBuildDistribution)
# Sample and Save Trajectory
trajectoriesWithIntentionDists = []
for trajectoryId in range(self.numTrajectories):
# Intention Prior For inference
otherWolfPossibleIntentionSpaces = {0: [(0, (1, 2))], 1: [(0, ())]}
otherIntentionType = np.random.choice(
[1, 0], p=[otherCompeteRate, 1 - otherCompeteRate])
otherWolfIntentionSpace = otherWolfPossibleIntentionSpaces[
otherIntentionType]
selfPossibleIntentionSpaces = {
0: [(0, (1, 2))],
0.5: [(0, (1, 2)), (0, ())],
1: [(0, ())]
}
selfWolfIntentionSpace = selfPossibleIntentionSpaces[
competeDetectionRate]
intentionSpacesForAllWolves = [
selfWolfIntentionSpace, otherWolfIntentionSpace
]
wolvesIntentionPriors = [{
tuple(intention): 1 / len(allPossibleIntentionsOneWolf)
for intention in allPossibleIntentionsOneWolf
} for allPossibleIntentionsOneWolf in intentionSpacesForAllWolves]
# Infer and update Intention
variablesForAllWolves = [[
intentionSpace
] for intentionSpace in intentionSpacesForAllWolves]
jointHypothesisSpaces = [
pd.MultiIndex.from_product(variables, names=['intention'])
for variables in variablesForAllWolves
]
concernedHypothesisVariable = ['intention']
priorDecayRate = 1
softPrior = SoftDistribution(priorDecayRate)
inferIntentionOneStepList = [
InferOneStep(jointHypothesisSpace, concernedHypothesisVariable,
calJointLikelihood, softPrior)
for jointHypothesisSpace in jointHypothesisSpaces
]
chooseIntention = sampleFromDistribution
valuePriorEndTime = -100
updateIntentions = [
UpdateIntention(intentionPrior, valuePriorEndTime,
adjustIntentionPriorGivenValueOfState,
perceptAction, inferIntentionOneStep,
chooseIntention)
for intentionPrior, inferIntentionOneStep in zip(
wolvesIntentionPriors, inferIntentionOneStepList)
]
# reset intention and adjuste intention prior attributes tools for multiple trajectory
intentionResetAttributes = [
'timeStep', 'lastState', 'lastAction', 'intentionPrior',
'formerIntentionPriors'
]
intentionResetAttributeValues = [
dict(
zip(intentionResetAttributes,
[0, None, None, intentionPrior, [intentionPrior]]))
for intentionPrior in wolvesIntentionPriors
]
resetIntentions = ResetObjects(intentionResetAttributeValues,
updateIntentions)
returnAttributes = ['formerIntentionPriors']
getIntentionDistributions = GetObjectsValuesOfAttributes(
returnAttributes, updateIntentions)
attributesToRecord = ['lastAction']
recordActionForUpdateIntention = RecordValuesForObjects(
attributesToRecord, updateIntentions)
# Wovels Generate Action
softPolicyInPlanning = SoftDistribution(softParameterInPlanning)
policyForCommittedAgentInPlanning = PolicyForCommittedAgent(
centralControlPolicyListBasedOnNumAgentsInWe,
softPolicyInPlanning, getStateOrActionThirdPersonPerspective)
policyForUncommittedAgentInPlanning = PolicyForUncommittedAgent(
possibleWolvesIds, heatSeekingDiscreteStochasticPolicy,
softPolicyInPlanning, getStateOrActionFirstPersonPerspective)
wolfChooseActionMethod = sampleFromDistribution
getSelfActionThirdPersonPerspective = lambda weIds, selfId: list(
weIds).index(selfId)
chooseCommittedAction = GetActionFromJointActionDistribution(
wolfChooseActionMethod, getSelfActionThirdPersonPerspective)
chooseUncommittedAction = sampleFromDistribution
wolvesSampleIndividualActionGivenIntentionList = [
SampleIndividualActionGivenIntention(
selfId, policyForCommittedAgentInPlanning,
policyForUncommittedAgentInPlanning, chooseCommittedAction,
chooseUncommittedAction) for selfId in possibleWolvesIds
]
wolvesSampleActions = [
SampleActionOnChangableIntention(
updateIntention,
wolvesSampleIndividualActionGivenIntention)
for updateIntention, wolvesSampleIndividualActionGivenIntention
in zip(updateIntentions,
wolvesSampleIndividualActionGivenIntentionList)
]
allIndividualSampleActions = sheepSampleActions + wolvesSampleActions
sampleActionMultiAgent = SampleActionMultiagent(
allIndividualSampleActions, recordActionForUpdateIntention)
trajectory = sampleTrajectory(sampleActionMultiAgent)
intentionDistributions = getIntentionDistributions()
trajectoryWithIntentionDists = [
tuple(list(SASRPair) + list(intentionDist)) for SASRPair,
intentionDist in zip(trajectory, intentionDistributions)
]
trajectoriesWithIntentionDists.append(
tuple(trajectoryWithIntentionDists))
resetIntentions()
#print(intentionDistributions[-1], otherCompeteRate)
trajectoryFixedParameters = {
'sheepPolicySoft': softParameterInPlanningForSheep,
'wolfPolicySoft': softParameterInPlanning,
'maxRunningSteps': maxRunningSteps,
'competePolicy': 'heatseeking',
'NNNumSimulations': NNNumSimulations,
'heatseekingPrecesion': heatseekingPrecesion
}
self.saveTrajectoryByParameters(trajectoriesWithIntentionDists,
trajectoryFixedParameters, parameters)
print(np.mean([len(tra) for tra in trajectoriesWithIntentionDists]))
def main():
parametersForTrajectoryPath = json.loads(sys.argv[1])
startSampleIndex = int(sys.argv[2])
endSampleIndex = int(sys.argv[3])
# parametersForTrajectoryPath['sampleOneStepPerTraj']=1 #0
# parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex, endSampleIndex)
trainSteps = int(parametersForTrajectoryPath['trainSteps'])
depth = int(parametersForTrajectoryPath['depth'])
dataSize = int(parametersForTrajectoryPath['dataSize'])
# parametersForTrajectoryPath = {}
# depth = 5
# dataSize = 5000
# trainSteps = 50000
# startSampleIndex = 0
# endSampleIndex = 100
killzoneRadius = 25
numSimulations = 200
maxRunningSteps = 100
fixedParameters = {
'maxRunningSteps': maxRunningSteps,
'numSimulations': numSimulations,
'killzoneRadius': killzoneRadius
}
trajectorySaveExtension = '.pickle'
dirName = os.path.dirname(__file__)
trajectoriesSaveDirectory = os.path.join(
dirName, '..', '..', '..', 'data', 'evaluateSupervisedLearning',
'multiMCTSAgentResNetNoPhysicsCenterControl',
'evaluateCenterControlTrajByCondition')
if not os.path.exists(trajectoriesSaveDirectory):
os.makedirs(trajectoriesSaveDirectory)
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]
reset = Reset(xBoundary, yBoundary, numOfAgent)
getSheepXPos = GetAgentPosFromState(sheepId, xPosIndex)
getWolfOneXPos = GetAgentPosFromState(wolfOneId, xPosIndex)
getWolfTwoXPos = GetAgentPosFromState(wolfTwoId, xPosIndex)
isTerminalOne = IsTerminal(getWolfOneXPos, getSheepXPos,
killzoneRadius)
isTerminalTwo = IsTerminal(getWolfTwoXPos, getSheepXPos,
killzoneRadius)
isTerminal = lambda state: isTerminalOne(state) or isTerminalTwo(state)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(
xBoundary, yBoundary)
transit = TransiteForNoPhysics(stayInBoundaryByReflectVelocity)
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7), (-10, 0), (-7, -7),
(0, -10), (7, -7), (0, 0)]
preyPowerRatio = 3
sheepActionSpace = list(
map(tuple,
np.array(actionSpace) * preyPowerRatio))
predatorPowerRatio = 2
wolfActionOneSpace = list(
map(tuple,
np.array(actionSpace) * predatorPowerRatio))
wolfActionTwoSpace = list(
map(tuple,
np.array(actionSpace) * predatorPowerRatio))
wolvesActionSpace = list(
it.product(wolfActionOneSpace, wolfActionTwoSpace))
# neural network init
numStateSpace = 6
numSheepActionSpace = len(sheepActionSpace)
numWolvesActionSpace = len(wolvesActionSpace)
regularizationFactor = 1e-4
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
generateSheepModel = GenerateModel(numStateSpace, numSheepActionSpace,
regularizationFactor)
# load save dir
NNModelSaveExtension = ''
NNModelSaveDirectory = os.path.join(
dirName, '..', '..', '..', 'data',
'evaluateEscapeMultiChasingNoPhysics',
'trainedResNNModelsMultiStillAction')
NNModelFixedParameters = {
'agentId': 0,
'maxRunningSteps': 150,
'numSimulations': 200,
'miniBatchSize': 256,
'learningRate': 0.0001
}
getNNModelSavePath = GetSavePath(NNModelSaveDirectory,
NNModelSaveExtension,
NNModelFixedParameters)
if not os.path.exists(NNModelSaveDirectory):
os.makedirs(NNModelSaveDirectory)
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepNNModel = generateSheepModel(sharedWidths * 5,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod,
dropoutRate)
sheepTrainedModelPath = getNNModelSavePath({
'trainSteps': 50000,
'depth': 5
})
sheepTrainedModel = restoreVariables(initSheepNNModel,
sheepTrainedModelPath)
sheepPolicy = ApproximatePolicy(sheepTrainedModel, sheepActionSpace)
generateWolvesModel = GenerateModel(numStateSpace,
numWolvesActionSpace,
regularizationFactor)
initWolvesNNModel = generateWolvesModel(sharedWidths * depth,
actionLayerWidths,
valueLayerWidths, resBlockSize,
initializationMethod,
dropoutRate)
NNModelSaveDirectory = os.path.join(
dirName, '..', '..', '..', 'data', 'evaluateSupervisedLearning',
'multiMCTSAgentResNetNoPhysicsCenterControl', 'trainedResNNModels')
wolfId = 1
NNModelFixedParametersWolves = {
'agentId': wolfId,
'maxRunningSteps': maxRunningSteps,
'numSimulations': numSimulations,
'miniBatchSize': 256,
'learningRate': 0.0001,
}
getNNModelSavePath = GetSavePath(NNModelSaveDirectory,
NNModelSaveExtension,
NNModelFixedParametersWolves)
wolvesTrainedModelPath = getNNModelSavePath({
'trainSteps': trainSteps,
'depth': depth,
'dataSize': dataSize
})
wolvesTrainedModel = restoreVariables(initWolvesNNModel,
wolvesTrainedModelPath)
wolfPolicy = ApproximatePolicy(wolvesTrainedModel, wolvesActionSpace)
from exec.evaluateNoPhysicsEnvWithRender import Render
import pygame as pg
from pygame.color import THECOLORS
screenColor = THECOLORS['black']
circleColorList = [
THECOLORS['green'], THECOLORS['red'], THECOLORS['orange']
]
circleSize = 10
saveImage = False
saveImageDir = os.path.join(dirName, '..', '..', '..', 'data',
'demoImg')
if not os.path.exists(saveImageDir):
os.makedirs(saveImageDir)
renderOn = False
render = None
if renderOn:
screen = pg.display.set_mode([xBoundary[1], yBoundary[1]])
render = Render(numOfAgent, xPosIndex, screen, screenColor,
circleColorList, circleSize, saveImage,
saveImageDir)
chooseActionList = [chooseGreedyAction, chooseGreedyAction]
sampleTrajectory = SampleTrajectoryWithRender(maxRunningSteps, transit,
isTerminal, reset,
chooseActionList, render,
renderOn)
# All agents' policies
policy = lambda state: [sheepPolicy(state), wolfPolicy(state)]
trajectories = [
sampleTrajectory(policy)
for sampleIndex in range(startSampleIndex, endSampleIndex)
]
saveToPickle(trajectories, trajectorySavePath)
def main():
DEBUG = 1
renderOn = 1
if DEBUG:
parametersForTrajectoryPath = {}
startSampleIndex = 1
endSampleIndex = 2
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex, endSampleIndex)
iterationIndex = 2
numTrainStepEachIteration = 1
numTrajectoriesPerIteration = 1
else:
parametersForTrajectoryPath = json.loads(sys.argv[1])
startSampleIndex = int(sys.argv[2])
endSampleIndex = int(sys.argv[3])
parametersForTrajectoryPath['sampleIndex'] = (startSampleIndex, endSampleIndex)
iterationIndex = int(parametersForTrajectoryPath['iterationIndex'])
numTrainStepEachIteration = int(parametersForTrajectoryPath['numTrainStepEachIteration'])
numTrajectoriesPerIteration = int(parametersForTrajectoryPath['numTrajectoriesPerIteration'])
# check file exists or not
dirName = os.path.dirname(__file__)
trajectoriesSaveDirectory = os.path.join(dirName, '..', '..', 'data', 'iterTrain2wolves1sheepMADDPGEnv', 'trajectories')
if not os.path.exists(trajectoriesSaveDirectory):
os.makedirs(trajectoriesSaveDirectory)
trajectorySaveExtension = '.pickle'
maxRunningSteps = 50
numSimulations = 250
killzoneRadius = 50
numTree = 2
fixedParameters = {'maxRunningSteps': maxRunningSteps, 'numSimulations': numSimulations, 'killzoneRadius': killzoneRadius}
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)
terminalRewardList = [collisonRewardSheep,collisonRewardWolf]
rewardMultiAgents = [rewardSheep, rewardWolf]
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)
generateWolvesModel = GenerateModel(numStateSpace, numWolvesActionSpace, regularizationFactor)
generateModelList = [generateSheepModel, generateWolvesModel]
sheepDepth = 9
wolfDepth = 9
depthList = [sheepDepth, wolfDepth]
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
sheepId,wolvesId = [0,1]
trainableAgentIds = [sheepId, wolvesId]
multiAgentNNmodel = [generateModel(sharedWidths * depth, actionLayerWidths, valueLayerWidths, resBlockSize, initializationMethod, dropoutRate) for depth, generateModel in zip(depthList, generateModelList)]
otherAgentApproximatePolicy = [lambda NNmodel, : ApproximatePolicy(NNmodel, sheepActionSpace), lambda NNmodel, : ApproximatePolicy(NNmodel, wolvesActionSpace)]
# NNGuidedMCTS init
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
getApproximatePolicy = [lambda NNmodel, : ApproximatePolicy(NNmodel, sheepActionSpace), lambda NNmodel, : ApproximatePolicy(NNmodel, wolvesActionSpace)]
getApproximateValue = [lambda NNmodel: ApproximateValue(NNmodel), lambda NNmodel: ApproximateValue(NNmodel)]
def getStateFromNode(node): return list(node.id.values())[0]
chooseActionInMCTS = sampleFromDistribution
composeMultiAgentTransitInSingleAgentMCTS = ComposeMultiAgentTransitInSingleAgentMCTS(chooseActionInMCTS)
composeSingleAgentGuidedMCTS = ComposeSingleAgentGuidedMCTS(numTree, numSimulations, actionSpaceList, terminalRewardList, selectChild, isTerminal, transit, getStateFromNode, getApproximatePolicy, getApproximateValue, composeMultiAgentTransitInSingleAgentMCTS)
prepareMultiAgentPolicy = PrepareMultiAgentPolicy(composeSingleAgentGuidedMCTS, otherAgentApproximatePolicy, trainableAgentIds)
# load model
NNModelSaveExtension = ''
NNModelSaveDirectory = os.path.join(dirName, '..', '..', 'data', 'iterTrain2wolves1sheepMADDPGEnv', 'NNModelRes')
if not os.path.exists(NNModelSaveDirectory):
os.makedirs(NNModelSaveDirectory)
generateNNModelSavePath = GetSavePath(NNModelSaveDirectory, NNModelSaveExtension, fixedParameters)
for agentId in trainableAgentIds:
modelPath = generateNNModelSavePath({'iterationIndex': iterationIndex - 1, 'agentId': agentId, 'numTrajectoriesPerIteration': numTrajectoriesPerIteration, 'numTrainStepEachIteration': numTrainStepEachIteration})
restoredNNModel = restoreVariables(multiAgentNNmodel[agentId], modelPath)
multiAgentNNmodel[agentId] = restoredNNModel
multiAgentPolicy = prepareMultiAgentPolicy(multiAgentNNmodel)
chooseActionList = [maxFromDistribution, maxFromDistribution]
def sampleAction(state):
actionDists = multiAgentPolicy(state)
action = [chooseAction(actionDist) for actionDist, chooseAction in zip(actionDists, chooseActionList)]
return action
render = lambda state: None
forwardOneStep = ForwardMultiAgentsOneStep(transit, rewardMultiAgents)
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 __call__(self, parameters):
print(parameters)
numWolves = parameters['numWolves']
numSheep = parameters['numSheep']
softParamterForValue = parameters['valuePriorSoftMaxBeta']
valuePriorEndTime = parameters['valuePriorEndTime']
## MDP Env
# state is all multi agent state # action is all multi agent action
xBoundary = [0,600]
yBoundary = [0,600]
numOfAgent = numWolves + numSheep
reset = Reset(xBoundary, yBoundary, numOfAgent)
possibleSheepIds = list(range(numSheep))
possibleWolvesIds = list(range(numSheep, numSheep + numWolves))
getSheepStatesFromAll = lambda state: np.array(state)[possibleSheepIds]
getWolvesStatesFromAll = lambda state: np.array(state)[possibleWolvesIds]
killzoneRadius = 25
isTerminal = IsTerminal(killzoneRadius, getSheepStatesFromAll, getWolvesStatesFromAll)
stayInBoundaryByReflectVelocity = StayInBoundaryByReflectVelocity(xBoundary, yBoundary)
interpolateOneFrame = InterpolateOneFrame(stayInBoundaryByReflectVelocity)
numFramesToInterpolate = 5
transit = TransitWithTerminalCheckOfInterpolation(numFramesToInterpolate, interpolateOneFrame, isTerminal)
maxRunningSteps = 52
timeCost = 1/maxRunningSteps
terminalBonus = 1
rewardFunction = RewardFunctionByTerminal(timeCost, terminalBonus, isTerminal)
forwardOneStep = ForwardOneStep(transit, rewardFunction)
sampleTrajectory = SampleTrajectory(maxRunningSteps, isTerminal, reset, forwardOneStep)
## MDP Policy
# Sheep Part
# Sheep Policy Function
numSheepPolicyStateSpace = 2 * (numWolves + 1)
sheepActionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7),
(-10, 0), (-7, -7), (0, -10), (7, -7), (0, 0)]
preyPowerRatio = 12
sheepIndividualActionSpace = list(map(tuple, np.array(sheepActionSpace) * preyPowerRatio))
numSheepActionSpace = len(sheepIndividualActionSpace)
regularizationFactor = 1e-4
generateSheepModel = GenerateModel(numSheepPolicyStateSpace, numSheepActionSpace, regularizationFactor)
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
sheepNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initSheepModel = generateSheepModel(sharedWidths * sheepNNDepth, actionLayerWidths, valueLayerWidths,
resBlockSize, initializationMethod, dropoutRate)
sheepModelPath = os.path.join('..', '..', 'data', 'preTrainModel',
'agentId=0.'+str(numWolves)+'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations=110_trainSteps=50000')
sheepNNModel = restoreVariables(initSheepModel, sheepModelPath)
sheepPolicy = ApproximatePolicy(sheepNNModel, sheepIndividualActionSpace)
# Sheep Generate Action
softParameterInPlanningForSheep = 2.5
softPolicyInPlanningForSheep = SoftDistribution(softParameterInPlanningForSheep)
softenSheepPolicy = lambda relativeAgentsStatesForSheepPolicy: softPolicyInPlanningForSheep(sheepPolicy(relativeAgentsStatesForSheepPolicy))
sheepChooseActionMethod = sampleFromDistribution
sheepSampleActions = [SampleActionOnFixedIntention(selfId, possibleWolvesIds, softenSheepPolicy, sheepChooseActionMethod) for selfId in possibleSheepIds]
# Wolves Part
# Policy Likelihood function: Wolf Centrol Control NN Policy Given Intention
numWolvesStateSpaces = [2 * (numInWe + numSheep)
for numInWe in range(2, numWolves + 1)]
actionSpace = [(10, 0), (7, 7), (0, 10), (-7, 7),
(-10, 0), (-7, -7), (0, -10), (7, -7), (0, 0)]
predatorPowerRatio = 8
wolfIndividualActionSpace = list(map(tuple, np.array(actionSpace) * predatorPowerRatio))
wolvesCentralControlActionSpaces = [list(it.product(wolfIndividualActionSpace, repeat = numInWe))
for numInWe in range(2, numWolves + 1)]
numWolvesCentralControlActionSpaces = [len(wolvesCentralControlActionSpace)
for wolvesCentralControlActionSpace in wolvesCentralControlActionSpaces]
regularizationFactor = 1e-4
generateWolvesCentralControlModels = [GenerateModel(numStateSpace, numActionSpace, regularizationFactor)
for numStateSpace, numActionSpace in zip(numWolvesStateSpaces, numWolvesCentralControlActionSpaces)]
sharedWidths = [128]
actionLayerWidths = [128]
valueLayerWidths = [128]
wolfNNDepth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initWolvesCentralControlModels = [generateWolvesCentralControlModel(sharedWidths * wolfNNDepth, actionLayerWidths, valueLayerWidths,
resBlockSize, initializationMethod, dropoutRate) for generateWolvesCentralControlModel in generateWolvesCentralControlModels]
NNNumSimulations = 250
wolvesModelPaths = [os.path.join('..', '..', 'data', 'preTrainModel',
'agentId=.'+str(len(actionSpace) * np.sum([10**_ for _ in
range(numInWe)]))+'_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations='+str(NNNumSimulations)+'_trainSteps=50000')
for numInWe in range(2, numWolves + 1)]
print(wolvesModelPaths)
wolvesCentralControlNNModels = [restoreVariables(initWolvesCentralControlModel, wolvesModelPath)
for initWolvesCentralControlModel, wolvesModelPath in zip(initWolvesCentralControlModels, wolvesModelPaths)]
wolvesCentralControlPolicies = [ApproximatePolicy(NNModel, actionSpace)
for NNModel, actionSpace in zip(wolvesCentralControlNNModels, wolvesCentralControlActionSpaces)]
# Wovels Generate Action
softParameterInPlanning = 2.5
softPolicyInPlanning = SoftDistribution(softParameterInPlanning)
wolvesPolicy = lambda state: wolvesCentralControlPolicies[numWolves - 2](state)
wolfChooseActionMethod = sampleFromDistribution
wolvesSampleAction = lambda state: wolfChooseActionMethod(softPolicyInPlanning(wolvesPolicy(state)))
def sampleAction(state):
action = list(wolvesSampleAction(state)) + [sheepSampleAction(state) for sheepSampleAction in sheepSampleActions]
return action
# Sample and Save Trajectory
trajectories = [sampleTrajectory(sampleAction) for _ in range(self.numTrajectories)]
wolfType = 'sharedReward'
trajectoryFixedParameters = {'sheepPolicySoft': softParameterInPlanningForSheep, 'wolfPolicySoft': softParameterInPlanning,
'maxRunningSteps': maxRunningSteps, 'hierarchy': 0, 'NNNumSimulations':NNNumSimulations, 'wolfType': wolfType}
self.saveTrajectoryByParameters(trajectories, trajectoryFixedParameters, parameters)
print(np.mean([len(tra) for tra in trajectories]))
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)
def main():
startTime = time.time()
DEBUG = 0
renderOn = 0
if DEBUG:
parametersForTrajectoryPath = {}
startSampleIndex = 5
endSampleIndex = 8
agentId = 0
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', 'trainSheepWithPretrrainWolves', '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]
generateWolvesModel = GenerateModel(numStateSpace, numWolvesActionSpace, regularizationFactor)
# wolf NN Policy
NNModelSaveExtension = ''
wolfTrainedModelPath = os.path.join(dirName, '..', '..', '..', '..', 'data', 'MADDPG2wolves1sheep', 'trainWolvesTwoCenterControlAction', 'trainedResNNModels', 'agentId=1_depth=9_learningRate=0.0001_maxRunningSteps=50_miniBatchSize=256_numSimulations=250_trainSteps=50000')
depth = 9
resBlockSize = 2
dropoutRate = 0.0
initializationMethod = 'uniform'
initWolfNNModel = generateWolvesModel(sharedWidths * depth, actionLayerWidths, valueLayerWidths, resBlockSize, initializationMethod, dropoutRate)
wolfTrainedModel = restoreVariables(initWolfNNModel, wolfTrainedModelPath)
wolfPolicy = ApproximatePolicy(wolfTrainedModel, wolvesActionSpace)
# MCTS
cInit = 1
cBase = 100
calculateScore = ScoreChild(cInit, cBase)
selectChild = SelectChild(calculateScore)
# prior
getActionPrior = lambda state: {action: 1 / len(sheepActionSpace) for action in sheepActionSpace}
# load chase nn policy
chooseActionInMCTS = sampleFromDistribution
def sheepTransit(state, action): return transit(
state, [action, chooseActionInMCTS(wolfPolicy(state))])
# initialize children; expand
initializeChildren = InitializeChildren(
sheepActionSpace, sheepTransit, getActionPrior)
isTerminal = lambda state: False
expand = Expand(isTerminal, initializeChildren)
# random rollout policy
def rolloutPolicy(
state): return [sheepActionSpace[np.random.choice(range(numSheepActionSpace))],sampleFromDistribution(wolfPolicy(state))]
rolloutHeuristic = lambda state: 0
maxRolloutSteps = 15
rollout = RollOut(rolloutPolicy, maxRolloutSteps, transit, rewardSheep, isTerminal, rolloutHeuristic)
sheepPolicy = 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 = lambda state: None
forwardOneStep = ForwardOneStep(transit, rewardSheep)
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()