def main():
independentVariables = OrderedDict()
# independentVariables['bufferSize'] = [1000, 5000, 10000, 20000]
# independentVariables['layerWidth'] = [[32], [32, 32], [64, 64, 64], [128, 128, 128, 128]]
# independentVariables['varianceDiscount'] = [0.995, 0.9995, 0.99995]
independentVariables['bufferSize'] = [1000, 5000]
independentVariables['layerWidth'] = [(32, 32), (32, 32, 32)]
independentVariables['varianceDiscount'] = [0.995, 0.9995]
modelSaveDirectory = "../trainedDDPGModels"
modelSaveExtension = '.ckpt'
getSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension)
evaluate = EvaluateNoiseAndMemorySize(getSavePath, saveModel=False)
levelNames = list(independentVariables.keys())
levelValues = list(independentVariables.values())
levelIndex = pd.MultiIndex.from_product(levelValues, names=levelNames)
toSplitFrame = pd.DataFrame(index=levelIndex)
resultDF = toSplitFrame.groupby(levelNames).apply(evaluate)
resultPath = os.path.join(
dirName, '..', '..', 'plots',
str(list(independentVariables.keys())) + '.pickle')
saveToPickle(resultDF, resultPath)
nCols = len(independentVariables['bufferSize'])
nRows = len(independentVariables['layerWidth'])
numplot = 1
axs = []
figure = plt.figure(dpi=200)
figure.set_size_inches(7.5, 6)
figure.suptitle('buffer: ' + str(independentVariables['bufferSize']) +
', layers: ' + str(independentVariables['layerWidth']) +
', varDiscount: ' +
str(independentVariables['varianceDiscount']),
fontsize=8)
for layerWidth, outterSubDf in resultDF.groupby('layerWidth'):
for bufferSize, innerSubDf in outterSubDf.groupby('bufferSize'):
subplot = figure.add_subplot(nRows, nCols, numplot)
axs.append(subplot)
numplot += 1
plt.ylim([-800, 400])
innerSubDf.T.plot(ax=subplot)
subplot.set_title('layerWidth = ' + str(layerWidth) +
' bufferSize = ' + str(bufferSize),
fontsize=5)
subplot.set_ylabel('MeanEpsReward', fontsize=5)
subplot.set_xlabel('Episode', fontsize=5)
subplot.tick_params(axis='both', which='major', labelsize=5)
subplot.tick_params(axis='both', which='minor', labelsize=5)
plt.legend(loc='best', prop={'size': 5})
plotPath = os.path.join(dirName, '..', '..', 'plots')
plt.savefig(os.path.join(plotPath, str(list(independentVariables.keys()))))
plt.show()
def __call__(self, env):
agent = Agent(
alpha=self.hyperparamDict['actorLR'],
beta=self.hyperparamDict['criticLR'],
input_dims=[env.observation_space.shape[0]],
tau=self.hyperparamDict['tau'],
env=env_norm(env) if self.hyperparamDict['normalizeEnv'] else env,
n_actions=env.action_space.shape[0],
units1=self.hyperparamDict['actorHiddenLayersWidths'],
units2=self.hyperparamDict['criticHiddenLayersWidths'],
actoractivationfunction=self.hyperparamDict['actorActivFunction'],
actorHiddenLayersWeightInit=self.
hyperparamDict['actorHiddenLayersWeightInit'],
actorHiddenLayersBiasInit=self.
hyperparamDict['actorHiddenLayersBiasInit'],
actorOutputWeightInit=self.hyperparamDict['actorOutputWeightInit'],
actorOutputBiasInit=self.hyperparamDict['actorOutputBiasInit'],
criticHiddenLayersWidths=self.
hyperparamDict['criticHiddenLayersWidths'],
criticActivFunction=self.hyperparamDict['criticActivFunction'],
criticHiddenLayersBiasInit=self.
hyperparamDict['criticHiddenLayersBiasInit'],
criticHiddenLayersWeightInit=self.
hyperparamDict['criticHiddenLayersWeightInit'],
criticOutputWeightInit=self.
hyperparamDict['criticOutputWeightInit'],
criticOutputBiasInit=self.hyperparamDict['criticOutputBiasInit'],
max_size=self.hyperparamDict['bufferSize'],
gamma=self.hyperparamDict['gamma'],
batch_size=self.hyperparamDict['minibatchSize'],
initnoisevar=self.hyperparamDict['noiseInitVariance'],
noiseDecay=self.hyperparamDict['varianceDiscount'],
noiseDacayStep=self.hyperparamDict['noiseDecayStartStep'],
minVar=self.hyperparamDict['minVar'],
path=self.hyperparamDict['modelSavePathPhil'])
episodeRewardList = []
meanEpsRewardList = []
for episode in range(1, self.hyperparamDict['maxEpisode'] + 1):
obs = env.reset()
epsReward = 0
for i in range(self.hyperparamDict['maxTimeStep']):
done = False
#env.render()
act = agent.choose_action(obs, env.action_space.low,
env.action_space.high)
new_state, reward, done, info = env.step(act)
agent.remember(obs, act, reward, new_state, int(done))
agent.learn()
epsReward += reward
obs = new_state
agent.runtime += 1
if done:
break
episodeRewardList.append(epsReward)
meanEpsRewardList.append(np.mean(episodeRewardList))
last100EpsMeanReward = np.mean(episodeRewardList[-100:])
if episode % 1 == 0:
print(
'episode: {}, last 100eps mean reward: {}, last eps reward: {} with {} steps'
.format(episode, last100EpsMeanReward, epsReward,
agent.runtime))
agent.save_models()
saveToPickle(meanEpsRewardList,
self.hyperparamDict['rewardSavePathPhil'])
return meanEpsRewardList
def simuliateOneParameter(parameterDict, evalNum, randomSeed, dt):
mujocoVisualize = False
demoVisualize = False
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
wolfColor = np.array([0.85, 0.35, 0.35])
sheepColor = np.array([0.35, 0.85, 0.35])
blockColor = np.array([0.25, 0.25, 0.25])
wolvesID = [0]
sheepsID = [1]
blocksID = [2]
numWolves = len(wolvesID)
numSheeps = len(sheepsID)
numBlocks = len(blocksID)
sheepMaxSpeed = 1.3
wolfMaxSpeed = 1.0
blockMaxSpeed = None
agentsMaxSpeedList = [wolfMaxSpeed] * numWolves + [sheepMaxSpeed
] * numSheeps
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [
blockSize
] * numBlocks
entityMaxSpeedList = [wolfMaxSpeed] * numWolves + [
sheepMaxSpeed
] * numSheeps + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList, isCollision)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList,
getPosFromAgentState, isCollision,
punishForOutOfBound)
rewardFunc = lambda state, action, nextState: \
list(rewardWolf(state, action, nextState)) + list(rewardSheep(state, action, nextState))
makePropertyList = MakePropertyList(transferNumberListToStr)
#changeCollisionReleventParameter
dmin = parameterDict['dmin']
dmax = parameterDict['dmax']
width = parameterDict['width']
midpoint = parameterDict['midpoint']
power = parameterDict['power']
timeconst = parameterDict['timeconst']
dampratio = parameterDict['dampratio']
geomIds = [1, 2, 3]
keyNameList = ['@solimp', '@solref']
valueList = [[[dmin, dmax, width, midpoint, power], [timeconst, dampratio]]
] * len(geomIds)
collisionParameter = makePropertyList(geomIds, keyNameList, valueList)
#changeSize
# geomIds=[1,2]
# keyNameList=['@size']
# valueList=[[[0.075,0.075]],[[0.1,0.1]]]
# sizeParameter=makePropertyList(geomIds,keyNameList,valueList)
#load env xml and change some geoms' property
physicsDynamicsPath = os.path.join(
dirName, '..', '..', 'environment', 'mujocoEnv',
'variousCollision_numBlocks=1_numSheeps=1_numWolves=1dt={}.xml'.format(
dt))
with open(physicsDynamicsPath) as f:
xml_string = f.read()
envXmlDict = xmltodict.parse(xml_string.strip())
envXmlPropertyDictList = [collisionParameter]
changeEnvXmlPropertFuntionyList = [changeGeomProperty]
for propertyDict, changeXmlProperty in zip(
envXmlPropertyDictList, changeEnvXmlPropertFuntionyList):
envXmlDict = changeXmlProperty(envXmlDict, propertyDict)
envXml = xmltodict.unparse(envXmlDict)
# print(envXmlDict['mujoco']['worldbody']['body'][0])
physicsModel = mujoco.load_model_from_xml(envXml)
physicsSimulation = mujoco.MjSim(physicsModel)
# MDP function
qPosInit = [0, 0] * numAgents
qVelInit = [0, 0] * numAgents
qVelInitNoise = 0
qPosInitNoise = 1
# reset=ResetUniformWithoutXPos(physicsSimulation, qPosInit, qVelInit, numAgents, numBlocks,qPosInitNoise, qVelInitNoise)
# fixReset=ResetFixWithoutXPos(physicsSimulation, qPosInit, qVelInit, numAgents,numBlocks)
# blocksState=[[0,-0.8,0,0]]
# reset=lambda :fixReset([-0.5,0.8,-0.5,0,0.5,0.8],[0,0,0,0,0,0],blocksState)
isTerminal = lambda state: False
numSimulationFrames = int(0.1 / dt)
print(numSimulationFrames)
# dt=0.01
damping = 2.5
reshapeAction = lambda action: action
# transit = TransitionFunctionWithoutXPos(physicsSimulation,numAgents , numSimulationFrames,damping*dt/numSimulationFrames,agentsMaxSpeedList,mujocoVisualize,reshapeAction)
maxRunningSteps = 10
# sampleTrajectory = SampleTrajectory(maxRunningSteps, transit, isTerminal, rewardFunc, reset)
# observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID, blocksID, getPosFromAgentState, getVelFromAgentState)
# observe = lambda state: [observeOneAgent(agentID)(state) for agentID in range(numAgents)]
# initObsForParams = observe(reset())
# obsShape = [initObsForParams[obsID].shape for obsID in range(len(initObsForParams))]
class ImpulsePolicy(object):
"""docstring for c"""
def __init__(self, initAction):
self.initAction = initAction
def __call__(self, state, timeStep):
action = [[0, 0], [0, 0]]
if timeStep == 0:
action = self.initAction
return action
worldDim = 2
trajList = []
startTime = time.time()
for i in range(evalNum):
np.random.seed(randomSeed + i)
initSpeedDirection = np.random.uniform(-np.pi / 2, np.pi / 2, 1)[0]
initSpeed = np.random.uniform(0, 1, 1)[0]
initActionDirection = np.random.uniform(-np.pi / 2, np.pi / 2, 1)[0]
initForce = np.random.uniform(0, 5, 1)[0]
# print(initSpeedDirection,initSpeed,initActionDirection,initForce)
fixReset = ResetFixWithoutXPos(physicsSimulation, qPosInit, qVelInit,
numAgents, numBlocks)
blocksState = [[0, 0, 0, 0]] #posX posY velX velY
reset = lambda: fixReset([-0.28, 0, 8, 8], [
initSpeed * np.cos(initSpeedDirection), initSpeed * np.sin(
initSpeedDirection), 0, 0
], blocksState)
transit = TransitionFunctionWithoutXPos(
physicsSimulation, numAgents, numSimulationFrames,
damping * dt * numSimulationFrames, agentsMaxSpeedList,
mujocoVisualize, isTerminal, reshapeAction)
initAction = [[
initForce * np.cos(initActionDirection),
initForce * np.sin(initActionDirection)
], [0, 0]]
impulsePolicy = ImpulsePolicy(initAction)
sampleTrajectory = SampleTrajectory(maxRunningSteps, transit,
isTerminal, rewardFunc, reset)
traj = sampleTrajectory(impulsePolicy)
# print('traj',traj[0])
trajList.append(traj)
# saveTraj
# print(trajList)
saveTraj = True
if saveTraj:
# trajSavePath = os.path.join(dirName,'traj', 'evaluateCollision', 'CollisionMoveTransitDamplingCylinder.pickle')
trajectorySaveExtension = '.pickle'
fixedParameters = {
'isMujoco': 1,
'isCylinder': 1,
'randomSeed': randomSeed,
'evalNum': evalNum
}
trajectoriesSaveDirectory = trajSavePath = os.path.join(
dirName, '..', 'trajectory', 'variousCollsiondt={}'.format(dt))
generateTrajectorySavePath = GetSavePath(trajectoriesSaveDirectory,
trajectorySaveExtension,
fixedParameters)
trajectorySavePath = generateTrajectorySavePath(parameterDict)
saveToPickle(trajList, trajectorySavePath)
# visualize
if demoVisualize:
entitiesColorList = [wolfColor] * numWolves + [
sheepColor
] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents,
getPosFromAgentState)
trajToRender = np.concatenate(trajList)
render(trajToRender)
endTime = time.time()
print("Time taken {} seconds to generate {} trajectories".format(
(endTime - startTime), evalNum))
def generateSingleCondition(condition):
debug = 0
if debug:
damping=2.0
frictionloss=0.0
masterForce=1.0
numWolves = 1
numSheeps = 1
numMasters = 1
maxTimeStep = 25
saveTraj=False
saveImage=True
visualizeMujoco=False
visualizeTraj = True
makeVideo=True
else:
# print(sys.argv)
# condition = json.loads(sys.argv[1])
damping = float(condition['damping'])
frictionloss = float(condition['frictionloss'])
masterForce = float(condition['masterForce'])
numWolves = 1
numSheeps = 1
numMasters = 1
maxTimeStep = 25
saveTraj=False
saveImage=True
visualizeMujoco=False
visualizeTraj = True
makeVideo=False
print("maddpg: , saveTraj: {}, visualize: {},damping; {},frictionloss: {}".format( str(saveTraj), str(visualizeMujoco),damping,frictionloss))
numAgents = numWolves + numSheeps+numMasters
numEntities = numAgents + numMasters
wolvesID = [0]
sheepsID = [1]
masterID = [2]
wolfSize = 0.075
sheepSize = 0.05
masterSize = 0.075
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [masterSize] * numMasters
wolfMaxSpeed = 1.0
blockMaxSpeed = None
entitiesMovableList = [True] * numAgents + [False] * numMasters
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList, getPosFromAgentState, isCollision,punishForOutOfBound)
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList, isCollision)
rewardMaster= lambda state, action, nextState: [-reward for reward in rewardWolf(state, action, nextState)]
rewardFunc = lambda state, action, nextState: \
list(rewardWolf(state, action, nextState)) + list(rewardSheep(state, action, nextState))+list(rewardMaster(state, action, nextState))
dirName = os.path.dirname(__file__)
# physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','rope','leased.xml')
makePropertyList=MakePropertyList(transferNumberListToStr)
geomIds=[1,2,3]
keyNameList=[0,1]
valueList=[[damping,damping]]*len(geomIds)
dampngParameter=makePropertyList(geomIds,keyNameList,valueList)
changeJointDampingProperty=lambda envDict,geomPropertyDict:changeJointProperty(envDict,geomPropertyDict,'@damping')
geomIds=[1,2,3]
keyNameList=[0,1]
valueList=[[frictionloss,frictionloss]]*len(geomIds)
frictionlossParameter=makePropertyList(geomIds,keyNameList,valueList)
changeJointFrictionlossProperty=lambda envDict,geomPropertyDict:changeJointProperty(envDict,geomPropertyDict,'@frictionloss')
physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','rope','leasedNew.xml')
# physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','rope','leased.xml')
with open(physicsDynamicsPath) as f:
xml_string = f.read()
envXmlDict = xmltodict.parse(xml_string.strip())
envXmlDict = xmltodict.parse(xml_string.strip())
envXmlPropertyDictList=[dampngParameter,frictionlossParameter]
changeEnvXmlPropertFuntionyList=[changeJointDampingProperty,changeJointFrictionlossProperty]
for propertyDict,changeXmlProperty in zip(envXmlPropertyDictList,changeEnvXmlPropertFuntionyList):
envXmlDict=changeXmlProperty(envXmlDict,propertyDict)
envXml=xmltodict.unparse(envXmlDict)
physicsModel = mujoco.load_model_from_xml(envXml)
physicsSimulation = mujoco.MjSim(physicsModel)
# print(physicsSimulation.model.body_pos)
# print(dir(physicsSimulation.model))
# print(dir(physicsSimulation.data),physicsSimulation.dataphysicsSimulation.data)
# print(physicsSimulation.data.qpos,dir(physicsSimulation.data.qpos))
# print(physicsSimulation.data.qpos,dir(physicsSimulation.data.qpos))
qPosInit = (0, ) * 24
qVelInit = (0, ) * 24
qPosInitNoise = 0.4
qVelInitNoise = 0
numAgent = 2
tiedAgentId = [0, 2]
ropePartIndex = list(range(3, 12))
maxRopePartLength = 0.06
reset = ResetUniformWithoutXPosForLeashed(physicsSimulation, qPosInit, qVelInit, numAgent, tiedAgentId,ropePartIndex, maxRopePartLength, qPosInitNoise, qVelInitNoise)
numSimulationFrames=10
isTerminal= lambda state: False
reshapeActionList = [ReshapeAction(5),ReshapeAction(5),ReshapeAction(masterForce)]
transit=TransitionFunctionWithoutXPos(physicsSimulation, numSimulationFrames, visualizeMujoco,isTerminal, reshapeActionList)
# damping=2.5
# numSimulationFrames =int(0.1/dt)
# agentsMaxSpeedList = [wolfMaxSpeed]* numWolves + [sheepMaxSpeed] * numSheeps
# reshapeAction = ReshapeAction()
# isTerminal = lambda state: False
# transit = TransitionFunctionWithoutXPos(physicsSimulation,numAgents , numSimulationFrames,damping*dt*numSimulationFrames,agentsMaxSpeedList,visualizeMujoco,isTerminal,reshapeAction)
maxRunningStepsToSample = 100
sampleTrajectory = SampleTrajectory(maxRunningStepsToSample, transit, isTerminal, rewardFunc, reset)
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID, masterID, getPosFromAgentState, getVelFromAgentState)
observe = lambda state: [observeOneAgent(agentID)(state) for agentID in range(numAgents)]
initObsForParams = observe(reset())
obsShape = [initObsForParams[obsID].shape[0] for obsID in range(len(initObsForParams))]
worldDim = 2
actionDim = worldDim * 2 + 1
layerWidth = [128, 128]
# ------------ model ------------------------
buildMADDPGModels = BuildMADDPGModels(actionDim, numAgents, obsShape)
modelsList = [buildMADDPGModels(layerWidth, agentID) for agentID in range(numAgents)]
# individStr = 'individ' if individualRewardWolf else 'shared'
# fileName = "maddpg{}wolves{}sheep{}blocks{}episodes{}stepSheepSpeed{}{}_agent".format(
# numWolves, numSheeps, numMasters, maxEpisode, maxTimeStep, sheepSpeedMultiplier, individStr)
# wolvesModelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg', fileName + str(i) + '60000eps') for i in wolvesID]
# [restoreVariables(model, path) for model, path in zip(wolvesModel, wolvesModelPaths)]
#
# actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
# policy = lambda allAgentsStates: [actOneStepOneModel(model, observe(allAgentsStates)) for model in modelsList]
# modelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg_ExpEpsLengthAndSheepSpeed', fileName + str(i)) for i in
# range(numAgents)]
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG')
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','timeconst='+str(timeconst)+'dampratio='+str(dampratio))
modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPGLeasedFixedEnv2','damping={}_frictionloss={}_masterForce={}'.format(damping,frictionloss,masterForce))
fileName = "maddpg{}episodes{}step_agent".format(maxEpisode, maxTimeStep)
modelPaths = [os.path.join(modelFolder, fileName + str(i) + '60000eps') for i in range(numAgents)]
[restoreVariables(model, path) for model, path in zip(modelsList, modelPaths)]
actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
policy = lambda allAgentsStates: [actOneStepOneModel(model, observe(allAgentsStates)) for model in modelsList]
trajList = []
numTrajToSample = 5
for i in range(numTrajToSample):
np.random.seed(i)
traj = sampleTrajectory(policy)
trajList.append(list(traj))
# saveTraj
if saveTraj:
trajFileName = "maddpg{}wolves{}sheep{}blocks{}eps{}stepSheepSpeed{}{}Traj".format(numWolves, numSheeps, numMasters, maxEpisode, maxTimeStep, sheepSpeedMultiplier, individStr)
trajSavePath = os.path.join(dirName, '..', 'trajectory', trajFileName)
saveToPickle(trajList, trajSavePath)
# visualize
if visualizeTraj:
demoFolder = os.path.join(dirName, '..', 'demos', 'mujocoMADDPGLeasedFixedEnv2','damping={}_frictionloss={}_masterForce={}'.format(damping,frictionloss,masterForce))
if not os.path.exists(demoFolder):
os.makedirs(demoFolder)
entitiesColorList = [wolfColor] * numWolves + [sheepColor] * numSheeps + [masterColor] * numMasters
render = Render(entitiesSizeList, entitiesColorList, numAgents,demoFolder,saveImage, getPosFromAgentState)
# print(trajList[0][0],'!!!3',trajList[0][1])
trajToRender = np.concatenate(trajList)
render(trajToRender)
def simuliateOneParameter(parameterDict):
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
wolfColor = np.array([0.85, 0.35, 0.35])
sheepColor = np.array([0.35, 0.85, 0.35])
blockColor = np.array([0.25, 0.25, 0.25])
wolvesID = [0,1]
sheepsID = [2]
blocksID = [3]
numWolves = len(wolvesID)
numSheeps = len(sheepsID)
numBlocks = len(blocksID)
sheepMaxSpeed = 1.3
wolfMaxSpeed =1.0
blockMaxSpeed = None
agentsMaxSpeedList = [wolfMaxSpeed]* numWolves + [sheepMaxSpeed] * numSheeps
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
entitiesSizeList = [wolfSize]* numWolves + [sheepSize] * numSheeps + [blockSize]* numBlocks
entityMaxSpeedList = [wolfMaxSpeed]* numWolves + [sheepMaxSpeed] * numSheeps + [blockMaxSpeed]* numBlocks
entitiesMovableList = [True]* numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList, isCollision)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList, getPosFromAgentState, isCollision, punishForOutOfBound)
rewardFunc = lambda state, action, nextState: \
list(rewardWolf(state, action, nextState)) + list(rewardSheep(state, action, nextState))
makePropertyList=MakePropertyList(transferNumberListToStr)
#changeCollisionReleventParameter
dmin=parameterDict['dmin']
dmax=parameterDict['dmax']
width=parameterDict['width']
midpoint=parameterDict['midpoint']
power=parameterDict['power']
timeconst=parameterDict['timeconst']
dampratio=parameterDict['dampratio']
geomIds=[1,2]
keyNameList=['@solimp','@solref']
valueList=[[[dmin,dmax,width,midpoint,power],[timeconst,dampratio]]]*len(geomIds)
collisionParameter=makePropertyList(geomIds,keyNameList,valueList)
#changeSize
# geomIds=[1,2]
# keyNameList=['@size']
# valueList=[[[0.075,0.075]],[[0.1,0.1]]]
# sizeParameter=makePropertyList(geomIds,keyNameList,valueList)
#load env xml and change some geoms' property
# physicsDynamicsPath=os.path.join(dirName,'multiAgentcollision.xml')
physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','multiAgentcollision.xml')
with open(physicsDynamicsPath) as f:
xml_string = f.read()
envXmlDict = xmltodict.parse(xml_string.strip())
print(envXmlDict)
envXmlPropertyDictList=[collisionParameter]
changeEnvXmlPropertFuntionyList=[changeGeomProperty]
for propertyDict,changeXmlProperty in zip(envXmlPropertyDictList,changeEnvXmlPropertFuntionyList):
envXmlDict=changeXmlProperty(envXmlDict,propertyDict)
envXml=xmltodict.unparse(envXmlDict)
physicsModel = mujoco.load_model_from_xml(envXml)
physicsSimulation = mujoco.MjSim(physicsModel)
# MDP function
qPosInit = [0, 0]*numAgents
qVelInit = [0, 0]*numAgents
qVelInitNoise = 0
qPosInitNoise = 1
# reset=ResetUniformWithoutXPos(physicsSimulation, qPosInit, qVelInit, numAgents, numBlocks,qPosInitNoise, qVelInitNoise)
fixReset=ResetFixWithoutXPos(physicsSimulation, qPosInit, qVelInit, numAgents,numBlocks)
blocksState=[[0,0,0,0]]
reset=lambda :fixReset([-0.5,0.8,-0.5,0,0.5,0.8],[0,0,0,0,0,0],blocksState)
isTerminal = lambda state: False
numSimulationFrames = 1
visualize=False
# physicsViewer=None
dt=0.1
damping=2.5
transit = TransitionFunctionWithoutXPos(physicsSimulation,numAgents , numSimulationFrames,damping*dt/numSimulationFrames,agentsMaxSpeedList,visualize,isTerminal)
maxRunningSteps = 100
sampleTrajectory = SampleTrajectory(maxRunningSteps, transit, isTerminal, rewardFunc, reset)
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID, blocksID, getPosFromAgentState, getVelFromAgentState)
observe = lambda state: [observeOneAgent(agentID)(state) for agentID in range(numAgents)]
initObsForParams = observe(reset())
obsShape = [initObsForParams[obsID].shape for obsID in range(len(initObsForParams))]
worldDim = 2
evalNum=1
trajectorySaveExtension = '.pickle'
fixedParameters = {'isMujoco': 1,'isCylinder':1,'evalNum':evalNum}
trajectoriesSaveDirectory=trajSavePath = os.path.join(dirName,'..','trajectory')
generateTrajectorySavePath = GetSavePath(trajectoriesSaveDirectory, trajectorySaveExtension, fixedParameters)
trajectorySavePath = generateTrajectorySavePath(parameterDict)
if not os.path.isfile(trajectorySavePath):
trajList =list()
for i in range(evalNum):
# policy =lambda state: [[-3,0] for agent in range(numAgents)]
np.random.seed(i)
# policy =lambda state: [np.random.uniform(-5,5,2) for agent in range(numAgents)]sss
# policy =lambda state: [[0,1] for agent in range(numAgents)]
policy =lambda state: [[1,0] ]
# policy =lambda state: [[np.random.uniform(0,1,1),0] ,[np.random.uniform(-1,0,1),0] ]
traj = sampleTrajectory(policy)
# print(i,'traj',[state[1] for state in traj[:2]])
# print(traj)
trajList.append( traj)
# saveTraj
saveTraj = True
if saveTraj:
# trajSavePath = os.path.join(dirName,'traj', 'evaluateCollision', 'CollisionMoveTransitDamplingCylinder.pickle')
saveToPickle(trajList, trajectorySavePath)
# visualize
# physicsViewer.render()
visualize = True
if visualize:
entitiesColorList = [wolfColor] * numWolves + [sheepColor] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents, getPosFromAgentState)
render(trajList)
def main():
stateDim = env.observation_space.shape[0]
actionDim = env.action_space.shape[0]
actionHigh = env.action_space.high
actionLow = env.action_space.low
actionBound = (actionHigh - actionLow) / 2
buildActorModel = BuildActorModel(stateDim, actionDim, actionBound)
actorLayerWidths = [30]
actorWriter, actorModel = buildActorModel(actorLayerWidths)
buildCriticModel = BuildCriticModel(stateDim, actionDim)
criticLayerWidths = [30]
criticWriter, criticModel = buildCriticModel(criticLayerWidths)
trainCriticBySASRQ = TrainCriticBySASRQ(learningRateCritic, gamma,
criticWriter)
trainCritic = TrainCritic(actByPolicyTarget, evaluateCriticTarget,
trainCriticBySASRQ)
trainActorFromGradients = TrainActorFromGradients(learningRateActor,
actorWriter)
trainActorOneStep = TrainActorOneStep(actByPolicyTrain,
trainActorFromGradients,
getActionGradients)
trainActor = TrainActor(trainActorOneStep)
paramUpdateInterval = 1
updateParameters = UpdateParameters(paramUpdateInterval, tau)
modelList = [actorModel, criticModel]
actorModel, criticModel = resetTargetParamToTrainParam(modelList)
trainModels = TrainDDPGModels(updateParameters, trainActor, trainCritic,
actorModel, criticModel)
noiseInitVariance = 3
varianceDiscount = .9995
noiseDecayStartStep = bufferSize
getNoise = GetExponentialDecayGaussNoise(noiseInitVariance,
varianceDiscount,
noiseDecayStartStep)
actOneStepWithNoise = ActDDPGOneStep(actionLow, actionHigh,
actByPolicyTrain, actorModel,
getNoise)
learningStartBufferSize = minibatchSize
sampleFromMemory = SampleFromMemory(minibatchSize)
learnFromBuffer = LearnFromBuffer(learningStartBufferSize,
sampleFromMemory, trainModels)
transit = TransitGymPendulum()
getReward = RewardGymPendulum(angle_normalize)
sampleOneStep = SampleOneStep(transit, getReward)
runDDPGTimeStep = RunTimeStep(actOneStepWithNoise, sampleOneStep,
learnFromBuffer, observe)
reset = ResetGymPendulum(seed)
runEpisode = RunEpisode(reset, runDDPGTimeStep, maxTimeStep,
isTerminalGymPendulum)
dirName = os.path.dirname(__file__)
modelPath = os.path.join(dirName, '..', 'trainedDDPGModels', 'pendulum')
getTrainedModel = lambda: trainModels.actorModel
modelSaveRate = 50
saveModel = SaveModel(modelSaveRate, saveVariables, getTrainedModel,
modelPath)
ddpg = RunAlgorithm(runEpisode, maxEpisode, [saveModel])
replayBuffer = deque(maxlen=int(bufferSize))
meanRewardList, trajectory = ddpg(replayBuffer)
dirName = os.path.dirname(__file__)
trajectoryPath = os.path.join(dirName, '..', 'trajectory',
'pendulumTrajectory1.pickle')
saveToPickle(trajectory, trajectoryPath)
# plots& plot
showDemo = True
if showDemo:
visualize = VisualizeGymPendulum()
visualize(trajectory)
plotResult = True
if plotResult:
plt.plot(list(range(maxEpisode)), meanRewardList)
plt.show()
def main():
statedim = env.observation_space.shape[0]
actiondim = env.action_space.shape[0]
actionHigh = env.action_space.high
actionLow = env.action_space.low
actionBound = (actionHigh - actionLow)/2
replaybuffer = deque(maxlen=buffersize)
paramUpdateFrequency = 1
totalrewards = []
meanreward = []
totalreward = []
buildActorModel = BuildActorModel(statedim, actiondim, actionBound)
actorWriter, actorModel = buildActorModel(actornumberlayers)
buildCriticModel = BuildCriticModel(statedim, actiondim)
criticWriter, criticModel = buildCriticModel(criticnumberlayers)
trainCritic = TrainCritic(criticlearningRate, gamma, criticWriter)
trainActor = TrainActor(actorlearningRate, actorWriter)
updateParameters = UpdateParameters(tau,paramUpdateFrequency)
actorModel= ReplaceParameters(actorModel)
criticModel= ReplaceParameters(criticModel)
trainddpgModels = TrainDDPGModels(updateParameters, trainActor, trainCritic, actorModel, criticModel)
getnoise = GetNoise(noiseDecay,minVar,noiseDacayStep)
getnoiseaction = GetNoiseAction(actorModel,actionLow, actionHigh)
learn = Learn(buffersize,batchsize,trainddpgModels,actiondim)
runtime = 0
trajectory = []
noisevar = initnoisevar
for episode in range(EPISODE):
state = env.reset()
rewards = 0
for i in range(maxTimeStep):
env.render()
noise,noisevar = getnoise(runtime,noisevar)
noiseaction = getnoiseaction(state,noise)
nextstate,reward,done,info = env.step(noiseaction)
learn(replaybuffer,state, noiseaction, nextstate,reward)
trajectory.append((state, noiseaction, nextstate,reward))
rewards += reward
state = nextstate
runtime += 1
print(actionHigh,actionLow)
if i == maxTimeStep-1:
totalrewards.append(rewards)
totalreward.append(rewards)
print('episode: ',episode,'reward:',rewards, 'noisevar',noisevar)
if episode % 100 == 0:
meanreward.append(np.mean(totalreward))
print('episode: ',episode,'meanreward:',np.mean(totalreward))
totalreward = []
plt.plot(range(EPISODE),totalrewards)
plt.xlabel('episode')
plt.ylabel('rewards')
plt.show()
# save Model
modelIndex = 0
actorFixedParam = {'actorModel': modelIndex}
criticFixedParam = {'criticModel': modelIndex}
parameters = {'env': Env_name, 'Eps': EPISODE, 'batchsize': batchsize,'buffersize': buffersize,'maxTimeStep':maxTimeStep,
'gamma': gamma, 'actorlearningRate': actorlearningRate, 'criticlearningRate': criticlearningRate,
'tau': tau, 'noiseDecay': noiseDecay, 'minVar': minVar, 'initnoisevar': initnoisevar}
modelSaveDirectory = "/path/to/logs/trainedDDPGModels"
modelSaveExtension = '.ckpt'
getSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension, actorFixedParam)
getSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension, criticFixedParam)
savePathDQN = getSavePath(parameters)
with actorModel.as_default():
saveVariables(actorModel, savePathDQN)
with criticModel.as_default():
saveVariables(criticModel, savePathDQN)
dirName = os.path.dirname(__file__)
trajectoryPath = os.path.join(dirName,'trajectory', 'HopperTrajectory.pickle')
saveToPickle(trajectory, trajectoryPath)
def main():
debug = 1
if debug:
numWolves = 1
numSheeps = 1
numBlocks = 1
saveTraj = False
visualizeTraj = True
maxTimeStep = 25
sheepSpeedMultiplier = 1
individualRewardWolf = 0
timeconst=0.5
dampratio=0.2
else:
print(sys.argv)
condition = json.loads(sys.argv[1])
numWolves = int(condition['numWolves'])
numSheeps = int(condition['numSheeps'])
numBlocks = int(condition['numBlocks'])
saveTraj = True
visualizeTraj = False
maxTimeStep = 50
sheepSpeedMultiplier = 1.25
individualRewardWolf = 1
print("maddpg: {} wolves, {} sheep, {} blocks, saveTraj: {}, visualize: {}".format(numWolves, numSheeps, numBlocks, str(saveTraj), str(visualizeTraj)))
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
wolvesID = list(range(numWolves))
sheepsID = list(range(numWolves, numAgents))
blocksID = list(range(numAgents, numEntities))
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [blockSize] * numBlocks
wolfMaxSpeed = 1.0
blockMaxSpeed = None
sheepMaxSpeedOriginal = 1.3
sheepMaxSpeed = sheepMaxSpeedOriginal * sheepSpeedMultiplier
entityMaxSpeedList = [wolfMaxSpeed] * numWolves + [sheepMaxSpeed] * numSheeps + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList, getPosFromAgentState, isCollision,
punishForOutOfBound)
if individualRewardWolf:
rewardWolf = RewardWolfIndividual(wolvesID, sheepsID, entitiesSizeList, isCollision)
else:
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList, isCollision)
rewardFunc = lambda state, action, nextState: list(rewardWolf(state, action, nextState)) + list(rewardSheep(state, action, nextState))
reset = ResetMultiAgentChasing(numAgents, numBlocks)
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID, blocksID, getPosFromAgentState, getVelFromAgentState)
observe = lambda state: [observeOneAgent(agentID)(state) for agentID in range(numAgents)]
reshapeAction = ReshapeAction()
getCollisionForce = GetCollisionForce()
applyActionForce = ApplyActionForce(wolvesID, sheepsID, entitiesMovableList)
applyEnvironForce = ApplyEnvironForce(numEntities, entitiesMovableList, entitiesSizeList,
getCollisionForce, getPosFromAgentState)
integrateState = IntegrateState(numEntities, entitiesMovableList, massList,
entityMaxSpeedList, getVelFromAgentState, getPosFromAgentState)
transit = TransitMultiAgentChasing(numEntities, reshapeAction, applyActionForce, applyEnvironForce, integrateState)
isTerminal = lambda state: False
maxRunningStepsToSample = 100
sampleTrajectory = SampleTrajectory(maxRunningStepsToSample, transit, isTerminal, rewardFunc, reset)
initObsForParams = observe(reset())
obsShape = [initObsForParams[obsID].shape[0] for obsID in range(len(initObsForParams))]
worldDim = 2
actionDim = worldDim * 2 + 1
layerWidth = [128, 128]
# ------------ model ------------------------
buildMADDPGModels = BuildMADDPGModels(actionDim, numAgents, obsShape)
modelsList = [buildMADDPGModels(layerWidth, agentID) for agentID in range(numAgents)]
dirName = os.path.dirname(__file__)
# individStr = 'individ' if individualRewardWolf else 'shared'
# fileName = "maddpg{}wolves{}sheep{}blocks{}episodes{}stepSheepSpeed{}{}_agent".format(
# numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep, sheepSpeedMultiplier, individStr)
fileName = "maddpg{}wolves{}sheep{}blocks60000episodes25stepSheepSpeed1shared_agent".format(numWolves, numSheeps, numBlocks)
# wolvesModelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg', fileName + str(i) + '60000eps') for i in wolvesID]
# [restoreVariables(model, path) for model, path in zip(wolvesModel, wolvesModelPaths)]
#
# actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
# policy = lambda allAgentsStates: [actOneStepOneModel(model, observe(allAgentsStates)) for model in modelsList]
# modelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg_ExpEpsLengthAndSheepSpeed', fileName + str(i)) for i in
# range(numAgents)]
modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','timeconst='+str(timeconst)+'dampratio='+str(dampratio))
modelPaths = [os.path.join(modelFolder, fileName + str(i) + '60000eps') for i in range(numAgents)]
[restoreVariables(model, path) for model, path in zip(modelsList, modelPaths)]
actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
policy = lambda allAgentsStates: [actOneStepOneModel(model, observe(allAgentsStates)) for model in modelsList]
trajList = []
numTrajToSample = 50
for i in range(numTrajToSample):
traj = sampleTrajectory(policy)
trajList.append(list(traj))
# saveTraj
if saveTraj:
trajFileName = "maddpg{}wolves{}sheep{}blocks{}eps{}stepSheepSpeed{}{}Traj".format(numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep, sheepSpeedMultiplier, individStr)
trajSavePath = os.path.join(dirName, '..', 'trajectory', trajFileName)
saveToPickle(trajList, trajSavePath)
# visualize
if visualizeTraj:
entitiesColorList = [wolfColor] * numWolves + [sheepColor] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents, getPosFromAgentState)
trajToRender = np.concatenate(trajList)
render(trajToRender)
def __call__(self,env):
env = env_norm(env) if self.fixedParameters['normalizeEnv'] else env
actionHigh = env.action_space.high
actionLow = env.action_space.low
actionBound = (actionHigh - actionLow) / 2
stateDim = env.observation_space.shape[0]
actionDim = env.action_space.shape[0]
meanreward = []
trajectory = []
totalreward = []
totalrewards = []
episodereward = []
replaybuffer = deque(maxlen=int(self.fixedParameters['bufferSize']))
buildActorModel = BuildActorModel(stateDim, actionDim,actionBound ,
self.fixedParameters['actorHiddenLayersWeightInit'],self.fixedParameters['actorHiddenLayersBiasInit'],
self.fixedParameters['actorOutputWeightInit'], self.fixedParameters['actorOutputBiasInit'],self.fixedParameters['actorActivFunction'],self.fixedParameters['gradNormClipValue'],self.fixedParameters['normalizeEnv'])
actorModel = buildActorModel(self.fixedParameters['actorHiddenLayersWidths'])
buildCriticModel = BuildCriticModel(stateDim, actionDim,
self.fixedParameters['criticHiddenLayersWeightInit'],self.fixedParameters['criticHiddenLayersBiasInit'],
self.fixedParameters['criticOutputWeightInit'], self.fixedParameters['criticOutputBiasInit'],
self.fixedParameters['criticActivFunction'],self.fixedParameters['gradNormClipValue'],self.fixedParameters['normalizeEnv'])
criticModel = buildCriticModel(self.fixedParameters['criticHiddenLayersWidths'])
trainCritic = TrainCritic(self.fixedParameters['criticLR'], self.fixedParameters['gamma'])
trainActor = TrainActor(self.fixedParameters['actorLR'])
updateParameters = UpdateParameters(self.fixedParameters['tau'])
trainddpgModels = TrainDDPGModels(updateParameters, trainActor, trainCritic, actorModel,criticModel)
getnoise = GetNoise(self.fixedParameters['varianceDiscount'],self.fixedParameters['minVar'],self.fixedParameters['noiseDecayStartStep'],self.fixedParameters['noiseInitVariance'])
getnoiseaction = GetNoiseAction(actorModel,actionLow, actionHigh)
learn = Learn(self.fixedParameters['bufferSize'],self.fixedParameters['minibatchSize'],trainddpgModels)
actorModel= ReplaceParameters(actorModel)
criticModel= ReplaceParameters(criticModel)
state = env.reset()
replaybuffer = fillbuffer(3000,self.bufferfill,env,replaybuffer,state)
for episode in range(1,self.fixedParameters['maxEpisode']+1):
state = env.reset()
rewards = 0
for j in range(self.fixedParameters['maxTimeStep']):
env.render()
noise = getnoise(self.runstep)
noiseaction = getnoiseaction(state,noise)
nextstate,reward,done,info = env.step(noiseaction)
learn(replaybuffer,state, noiseaction, nextstate,reward)
trajectory.append((state, noiseaction, nextstate,reward))
rewards += reward
state = nextstate
self.runstep += 1
if j == self.fixedParameters['maxTimeStep']-1:
totalrewards.append(rewards)
totalreward.append(rewards)
print('episode: ',episode,'reward:',rewards,'runstep',self.runstep)
episodereward.append(np.mean(totalrewards))
print('epireward',np.mean(totalrewards))
if episode % 100 == 0:
meanreward.append(np.mean(totalreward))
print('episode: ',episode,'meanreward:',np.mean(totalreward))
totalreward = []
with actorModel.as_default():
saveVariables(actorModel, self.fixedParameters['modelSavePathMartin'])
with criticModel.as_default():
saveVariables(criticModel, self.fixedParameters['modelSavePathMartin'])
saveToPickle(meanreward, self.fixedParameters['rewardSavePathMartin'])
return episodereward
def main():
# wolvesID = [0, 1]
# sheepsID = [2]
# blocksID = [3]
visualize = False
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
sheepMaxSpeed = 1.3
wolfMaxSpeed = 1.0
blockMaxSpeed = None
wolfColor = np.array([0.85, 0.35, 0.35])
sheepColor = np.array([0.35, 0.85, 0.35])
blockColor = np.array([0.25, 0.25, 0.25])
wolvesID = [0]
sheepsID = [1]
blocksID = [2]
numWolves = len(wolvesID)
numSheeps = len(sheepsID)
numBlocks = len(blocksID)
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [
blockSize
] * numBlocks
entityMaxSpeedList = [wolfMaxSpeed] * numWolves + [
sheepMaxSpeed
] * numSheeps + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList, isCollision)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList,
getPosFromAgentState, isCollision,
punishForOutOfBound)
rewardFunc = lambda state, action, nextState: \
list(rewardWolf(state, action, nextState)) + list(rewardSheep(state, action, nextState))
reset = ResetMultiAgentChasing(numAgents, numBlocks)
# reset=lambda :np.array([[-0.5,0,0,0],[0.5,0,0,0]])
reset = lambda: np.array([[-0.5, 0, 0, 0], [8, 8, 0, 0], [0, -8, 0, 0]])
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID,
blocksID, getPosFromAgentState,
getVelFromAgentState)
observe = lambda state: [
observeOneAgent(agentID)(state) for agentID in range(numAgents)
]
# reshapeAction = ReshapeAction()
reshapeAction = lambda action: action
getCollisionForce = GetCollisionForce()
applyActionForce = ApplyActionForce(wolvesID, sheepsID,
entitiesMovableList)
applyEnvironForce = ApplyEnvironForce(numEntities, entitiesMovableList,
entitiesSizeList, getCollisionForce,
getPosFromAgentState)
integrateState = FixedIntegrateState(numEntities, entitiesMovableList,
massList, entityMaxSpeedList,
getVelFromAgentState,
getPosFromAgentState)
transit = TransitMultiAgentChasing(numEntities, reshapeAction,
applyActionForce, applyEnvironForce,
integrateState)
isTerminal = lambda state: False
maxRunningSteps = 100
sampleTrajectory = SampleTrajectory(maxRunningSteps, transit, isTerminal,
rewardFunc, reset)
initObsForParams = observe(reset())
obsShape = [
initObsForParams[obsID].shape for obsID in range(len(initObsForParams))
]
worldDim = 2
trajList = []
evalNum = 1
for i in range(evalNum):
# policy =lambda state: [[-3,0] for agent in range(numAgents)]
np.random.seed(i)
# policy =lambda state: [np.random.uniform(-5,5,2) for agent in range(numAgents)]
# policy =lambda state: [[1,1] for agent in range(numAgents)]
# policy =lambda state: [[float(np.random.uniform(0,1,1)),0] ,[float(np.random.uniform(-1,0,1)),0] ]
policy = lambda state: [[1, 0], [0, 0]]
traj = sampleTrajectory(policy)
# print(i,'traj',[state[1] for state in traj)
# print('traj',[tr[0][0][0] for tr in traj])
print('traj', traj[0])
trajList.append(traj)
# saveTraj
saveTraj = True
# print(trajList[0][0][0])
if saveTraj:
trajSavePath = os.path.join(dirName, '..', 'trajectory',
'LineMoveFixedBaseLine.pickle')
saveToPickle(trajList, trajSavePath)
# print('traj',trajList[25])
# visualize
if visualize:
entitiesColorList = [wolfColor] * numWolves + [
sheepColor
] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents,
getPosFromAgentState)
trajToRender = np.concatenate(trajList)
render(trajToRender)
def main():
debug = 1
if debug:
numWolves = 3
numSheeps = 1
numBlocks = 2
timeconst = 0.5
dampratio = 0.2
saveTraj = False
visualizeTraj = True
maxTimeStep = 25
sheepSpeedMultiplier = 1.0
individualRewardWolf = 0
hasWalls = 1.5
dt = 0.05
visualizeMujoco = True
else:
print(sys.argv)
condition = json.loads(sys.argv[1])
numWolves = int(condition['numWolves'])
numSheeps = int(condition['numSheeps'])
numBlocks = int(condition['numBlocks'])
saveTraj = True
visualizeTraj = False
maxTimeStep = 50
sheepSpeedMultiplier = 1.25
individualRewardWolf = 1
print(
"maddpg: {} wolves, {} sheep, {} blocks, saveTraj: {}, visualize: {}".
format(numWolves, numSheeps, numBlocks, str(saveTraj),
str(visualizeTraj)))
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
wolvesID = list(range(numWolves))
sheepsID = list(range(numWolves, numAgents))
blocksID = list(range(numAgents, numEntities))
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [
blockSize
] * numBlocks
wolfMaxSpeed = 1.0
blockMaxSpeed = None
sheepMaxSpeedOriginal = 1.3
sheepMaxSpeed = sheepMaxSpeedOriginal * sheepSpeedMultiplier
entityMaxSpeedList = [wolfMaxSpeed] * numWolves + [
sheepMaxSpeed
] * numSheeps + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList,
getPosFromAgentState, isCollision,
punishForOutOfBound)
if individualRewardWolf:
rewardWolf = RewardWolfIndividual(wolvesID, sheepsID, entitiesSizeList,
isCollision)
else:
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList,
isCollision)
rewardFunc = lambda state, action, nextState: list(
rewardWolf(state, action, nextState)) + list(
rewardSheep(state, action, nextState))
dirName = os.path.dirname(__file__)
# physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','numBlocks=1_numSheeps=1_numWolves=1.xml')
if hasWalls:
# physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','hasWalls=1_numBlocks={}_numSheeps={}_numWolves={}timeconst={}dampratio={}.xml'.format(numBlocks,numSheeps,numWolves,timeconst,dampratio))
physicsDynamicsPath = os.path.join(
dirName, '..', '..', 'environment', 'mujocoEnv',
'dt={}'.format(dt),
'hasWalls={}_numBlocks={}_numSheeps={}_numWolves={}timeconst={}dampratio={}.xml'
.format(hasWalls, numBlocks, numSheeps, numWolves, timeconst,
dampratio))
# physicsDynamicsPath=os.path.join(dirName,'..','..','environment','mujocoEnv','crossTest3w1s2b.xml')
else:
physicsDynamicsPath = os.path.join(
dirName, '..', '..', 'environment', 'mujocoEnv',
'numBlocks={}_numSheeps={}_numWolves={}timeconst={}dampratio={}.xml'
.format(numBlocks, numSheeps, numWolves, timeconst, dampratio))
with open(physicsDynamicsPath) as f:
xml_string = f.read()
envXmlDict = xmltodict.parse(xml_string.strip())
envXml = xmltodict.unparse(envXmlDict)
physicsModel = mujoco.load_model_from_xml(envXml)
physicsSimulation = mujoco.MjSim(physicsModel)
print(physicsSimulation.model.body_pos)
# print(dir(physicsSimulation.model))
# print(dir(physicsSimulation.data),physicsSimulation.dataphysicsSimulation.data)
# print(physicsSimulation.data.qpos,dir(physicsSimulation.data.qpos))
# print(physicsSimulation.data.qpos,dir(physicsSimulation.data.qpos))
qPosInit = [0, 0] * numAgents
qVelInit = [0, 0] * numAgents
qVelInitNoise = 0 * hasWalls
qPosInitNoise = 0.8 * hasWalls
getBlockRandomPos = lambda: np.random.uniform(-0.7 * hasWalls, +0.7 *
hasWalls, 2)
getBlockSpeed = lambda: np.zeros(2)
numQPos = len(physicsSimulation.data.qpos)
numQVel = len(physicsSimulation.data.qvel)
sampleAgentsQPos = lambda: np.asarray(qPosInit) + np.random.uniform(
low=-qPosInitNoise, high=qPosInitNoise, size=numQPos)
sampleAgentsQVel = lambda: np.asarray(qVelInit) + np.random.uniform(
low=-qVelInitNoise, high=qVelInitNoise, size=numQVel)
minDistance = 0.2 + 2 * blockSize
isOverlap = IsOverlap(minDistance)
sampleBlockState = SampleBlockState(numBlocks, getBlockRandomPos,
getBlockSpeed, isOverlap)
reset = ResetUniformWithoutXPos(physicsSimulation, numAgents, numBlocks,
sampleAgentsQPos, sampleAgentsQVel,
sampleBlockState)
damping = 2.5
numSimulationFrames = int(0.1 / dt)
agentsMaxSpeedList = [wolfMaxSpeed] * numWolves + [sheepMaxSpeed
] * numSheeps
reshapeAction = ReshapeAction()
isTerminal = lambda state: False
transit = TransitionFunctionWithoutXPos(physicsSimulation, numAgents,
numSimulationFrames,
damping * dt * numSimulationFrames,
agentsMaxSpeedList,
visualizeMujoco, isTerminal,
reshapeAction)
maxRunningStepsToSample = 100
sampleTrajectory = SampleTrajectory(maxRunningStepsToSample, transit,
isTerminal, rewardFunc, reset)
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID,
blocksID, getPosFromAgentState,
getVelFromAgentState)
observe = lambda state: [
observeOneAgent(agentID)(state) for agentID in range(numAgents)
]
initObsForParams = observe(reset())
obsShape = [
initObsForParams[obsID].shape[0]
for obsID in range(len(initObsForParams))
]
worldDim = 2
actionDim = worldDim * 2 + 1
layerWidth = [128, 128]
# ------------ model ------------------------
buildMADDPGModels = BuildMADDPGModels(actionDim, numAgents, obsShape)
modelsList = [
buildMADDPGModels(layerWidth, agentID) for agentID in range(numAgents)
]
# individStr = 'individ' if individualRewardWolf else 'shared'
# fileName = "maddpg{}wolves{}sheep{}blocks{}episodes{}stepSheepSpeed{}{}_agent".format(
# numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep, sheepSpeedMultiplier, individStr)
# wolvesModelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg', fileName + str(i) + '60000eps') for i in wolvesID]
# [restoreVariables(model, path) for model, path in zip(wolvesModel, wolvesModelPaths)]
#
# actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
# policy = lambda allAgentsStates: [actOneStepOneModel(model, observe(allAgentsStates)) for model in modelsList]
# modelPaths = [os.path.join(dirName, '..', 'trainedModels', '3wolvesMaddpg_ExpEpsLengthAndSheepSpeed', fileName + str(i)) for i in
# range(numAgents)]
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG')
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','timeconst='+str(timeconst)+'dampratio='+str(dampratio))
if hasWalls:
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','hasWalls=1'+'numBlocks='+str(numBlocks)+'numSheeps='+str(numSheeps)+'numWolves='+str(numWolves)+'timeconst='+str(timeconst)+'dampratio='+str(dampratio))
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','dt={}'.format(dt),'hasWalls='+str(hasWalls)+'numBlocks='+str(numBlocks)+'numSheeps='+str(numSheeps)+'numWolves='+str(numWolves)+'timeconst='+str(timeconst)+'dampratio='+str(dampratio)+'individualRewardWolf='+str(individualRewardWolf)+'sheepSpeedMultiplier='+str(sheepSpeedMultiplier))
# modelFolder = os.path.join(dirName, '..', 'trainedModels', 'mujocoMADDPG','dt={}'.format(dt),'hasWalls='+str(hasWalls)+'numBlocks='+str(numBlocks)+'numSheeps='+str(numSheeps)+'numWolves='+str(numWolves)+'timeconst='+str(timeconst)+'dampratio='+str(dampratio)+'individualRewardWolf='+str(individualRewardWolf)+'sheepSpeedMultiplier='+str(sheepSpeedMultiplier))
modelFolder = os.path.join(
dirName, '..', 'trainedModels', 'mujocoMADDPGeavluateWall',
'dt={}'.format(dt),
'hasWalls=' + str(hasWalls) + 'numBlocks=' + str(numBlocks) +
'numSheeps=' + str(numSheeps) + 'numWolves=' + str(numWolves) +
'timeconst=' + str(timeconst) + 'dampratio=' + str(dampratio) +
'individualRewardWolf=' + str(individualRewardWolf) +
'sheepSpeedMultiplier=' + str(sheepSpeedMultiplier))
individStr = 'individ' if individualRewardWolf else 'shared'
fileName = "maddpghasWalls={}{}wolves{}sheep{}blocks{}episodes{}stepSheepSpeed{}{}_agent".format(
hasWalls, numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep,
sheepSpeedMultiplier, individStr)
else:
modelFolder = os.path.join(
dirName, '..', 'trainedModels', 'mujocoMADDPG',
'numBlocks=' + str(numBlocks) + 'numSheeps=' + str(numSheeps) +
'numWolves=' + str(numWolves) + 'timeconst=' + str(timeconst) +
'dampratio=' + str(dampratio))
fileName = "maddpg{}wolves{}sheep{}blocks60000episodes25stepSheepSpeed1.0shared_agent".format(
numWolves, numSheeps, numBlocks)
modelPaths = [
os.path.join(modelFolder, fileName + str(i) + '60000eps')
for i in range(numAgents)
]
[
restoreVariables(model, path)
for model, path in zip(modelsList, modelPaths)
]
actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
policy = lambda allAgentsStates: [
actOneStepOneModel(model, observe(allAgentsStates))
for model in modelsList
]
trajList = []
numTrajToSample = 50
for i in range(numTrajToSample):
np.random.seed(i)
traj = sampleTrajectory(policy)
trajList.append(list(traj))
# saveTraj
if saveTraj:
trajFileName = "maddpg{}wolves{}sheep{}blocks{}eps{}stepSheepSpeed{}{}Traj".format(
numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep,
sheepSpeedMultiplier, individStr)
trajSavePath = os.path.join(dirName, '..', 'trajectory', trajFileName)
saveToPickle(trajList, trajSavePath)
# visualize
if visualizeTraj:
entitiesColorList = [wolfColor] * numWolves + [
sheepColor
] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents,
getPosFromAgentState)
print(trajList[0][0], '!!!3', trajList[0][1])
trajToRender = np.concatenate(trajList)
render(trajToRender)
def main():
stateDim = env.observation_space.shape[0]
actionDim = env.action_space.shape[0]
actionHigh = env.action_space.high
actionLow = env.action_space.low
actionBound = (actionHigh - actionLow) / 2
buildActorModel = BuildActorModel(stateDim, actionDim, actionBound)
actorLayerWidths = [30]
actorWriter, actorModel = buildActorModel(actorLayerWidths)
buildCriticModel = BuildCriticModel(stateDim, actionDim)
criticLayerWidths = [30]
criticWriter, criticModel = buildCriticModel(criticLayerWidths)
trainCriticBySASRQ = TrainCriticBySASRQ(learningRateCritic, gamma,
criticWriter)
trainCritic = TrainCritic(actByPolicyTarget, evaluateCriticTarget,
trainCriticBySASRQ)
trainActorFromGradients = TrainActorFromGradients(learningRateActor,
actorWriter)
trainActorOneStep = TrainActorOneStep(actByPolicyTrain,
trainActorFromGradients,
getActionGradients)
trainActor = TrainActor(trainActorOneStep)
paramUpdateInterval = 1
updateParameters = UpdateParameters(paramUpdateInterval, tau)
modelList = [actorModel, criticModel]
actorModel, criticModel = resetTargetParamToTrainParam(modelList)
trainModels = TrainDDPGModels(updateParameters, trainActor, trainCritic,
actorModel, criticModel)
noiseInitVariance = 1 # control exploration
varianceDiscount = .99995
noiseDecayStartStep = bufferSize
minVar = .1
getNoise = GetExponentialDecayGaussNoise(noiseInitVariance,
varianceDiscount,
noiseDecayStartStep, minVar)
actOneStepWithNoise = ActDDPGOneStep(actionLow, actionHigh,
actByPolicyTrain, actorModel,
getNoise)
learningStartBufferSize = minibatchSize
sampleFromMemory = SampleFromMemory(minibatchSize)
learnFromBuffer = LearnFromBuffer(learningStartBufferSize,
sampleFromMemory, trainModels)
transit = TransitGymMountCarContinuous()
isTerminal = IsTerminalMountCarContin()
getReward = RewardMountCarContin(isTerminal)
sampleOneStep = SampleOneStep(transit, getReward)
runDDPGTimeStep = RunTimeStep(actOneStepWithNoise, sampleOneStep,
learnFromBuffer)
resetLow = -1
resetHigh = 0.4
reset = ResetMountCarContin(seed=None)
runEpisode = RunEpisode(reset, runDDPGTimeStep, maxTimeStep, isTerminal)
ddpg = RunAlgorithm(runEpisode, maxEpisode)
replayBuffer = deque(maxlen=int(bufferSize))
meanRewardList, trajectory = ddpg(replayBuffer)
trainedActorModel, trainedCriticModel = trainModels.getTrainedModels()
# save Model
modelIndex = 0
actorFixedParam = {'actorModel': modelIndex}
criticFixedParam = {'criticModel': modelIndex}
parameters = {
'env': ENV_NAME,
'Eps': maxEpisode,
'timeStep': maxTimeStep,
'batch': minibatchSize,
'gam': gamma,
'lrActor': learningRateActor,
'lrCritic': learningRateCritic,
'noiseVar': noiseInitVariance,
'varDiscout': varianceDiscount,
'resetLow': resetLow,
'High': resetHigh
}
modelSaveDirectory = "../trainedDDPGModels"
modelSaveExtension = '.ckpt'
getActorSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension,
actorFixedParam)
getCriticSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension,
criticFixedParam)
savePathActor = getActorSavePath(parameters)
savePathCritic = getCriticSavePath(parameters)
with actorModel.as_default():
saveVariables(trainedActorModel, savePathActor)
with criticModel.as_default():
saveVariables(trainedCriticModel, savePathCritic)
dirName = os.path.dirname(__file__)
trajectoryPath = os.path.join(dirName, '..', 'trajectory',
'mountCarTrajectoryOriginalReset1.pickle')
saveToPickle(trajectory, trajectoryPath)
# plots& plot
showDemo = False
if showDemo:
visualize = VisualizeMountCarContin()
visualize(trajectory)
plotResult = True
if plotResult:
plt.plot(list(range(maxEpisode)), meanRewardList)
plt.show()
def simuliateOneParameter(parameterOneCondiiton, evalNum, randomSeed):
saveTraj = True
visualizeTraj = False
visualizeMujoco = False
numWolves = parameterOneCondiiton['numWolves']
numSheeps = parameterOneCondiiton['numSheeps']
numBlocks = parameterOneCondiiton['numBlocks']
timeconst = parameterOneCondiiton['timeconst']
dampratio = parameterOneCondiiton['dampratio']
maxTimeStep = parameterOneCondiiton['maxTimeStep']
sheepSpeedMultiplier = parameterOneCondiiton['sheepSpeedMultiplier']
individualRewardWolf = parameterOneCondiiton['individualRewardWolf']
hasWalls = parameterOneCondiiton['hasWalls']
dt = parameterOneCondiiton['dt']
print(
"maddpg: {} wolves, {} sheep, {} blocks, saveTraj: {}, visualize: {}".
format(numWolves, numSheeps, numBlocks, str(saveTraj),
str(visualizeTraj)))
numAgents = numWolves + numSheeps
numEntities = numAgents + numBlocks
wolvesID = list(range(numWolves))
sheepsID = list(range(numWolves, numAgents))
blocksID = list(range(numAgents, numEntities))
wolfSize = 0.075
sheepSize = 0.05
blockSize = 0.2
entitiesSizeList = [wolfSize] * numWolves + [sheepSize] * numSheeps + [
blockSize
] * numBlocks
wolfMaxSpeed = 1.0
blockMaxSpeed = None
sheepMaxSpeedOriginal = 1.3
sheepMaxSpeed = sheepMaxSpeedOriginal * sheepSpeedMultiplier
entityMaxSpeedList = [wolfMaxSpeed] * numWolves + [
sheepMaxSpeed
] * numSheeps + [blockMaxSpeed] * numBlocks
entitiesMovableList = [True] * numAgents + [False] * numBlocks
massList = [1.0] * numEntities
isCollision = IsCollision(getPosFromAgentState)
punishForOutOfBound = PunishForOutOfBound()
rewardSheep = RewardSheep(wolvesID, sheepsID, entitiesSizeList,
getPosFromAgentState, isCollision,
punishForOutOfBound)
if individualRewardWolf:
rewardWolf = RewardWolfIndividual(wolvesID, sheepsID, entitiesSizeList,
isCollision)
else:
rewardWolf = RewardWolf(wolvesID, sheepsID, entitiesSizeList,
isCollision)
rewardFunc = lambda state, action, nextState: list(
rewardWolf(state, action, nextState)) + list(
rewardSheep(state, action, nextState))
dirName = os.path.dirname(__file__)
if hasWalls:
physicsDynamicsPath = os.path.join(
dirName, '..', '..', 'environment', 'mujocoEnv',
'dt={}'.format(dt),
'hasWalls={}_numBlocks={}_numSheeps={}_numWolves={}timeconst={}dampratio={}.xml'
.format(hasWalls, numBlocks, numSheeps, numWolves, timeconst,
dampratio))
else:
physicsDynamicsPath = os.path.join(
dirName, '..', '..', 'environment', 'mujocoEnv',
'numBlocks={}_numSheeps={}_numWolves={}timeconst={}dampratio={}.xml'
.format(numBlocks, numSheeps, numWolves, timeconst, dampratio))
with open(physicsDynamicsPath) as f:
xml_string = f.read()
envXmlDict = xmltodict.parse(xml_string.strip())
envXml = xmltodict.unparse(envXmlDict)
physicsModel = mujoco.load_model_from_xml(envXml)
physicsSimulation = mujoco.MjSim(physicsModel)
# print(physicsSimulation.model.body_pos)
qPosInit = [0, 0] * numAgents
qVelInit = [0, 0] * numAgents
qVelInitNoise = 0 * hasWalls
qPosInitNoise = 0.8 * hasWalls
getBlockRandomPos = lambda: np.random.uniform(-0.7 * hasWalls, +0.7 *
hasWalls, 2)
getBlockSpeed = lambda: np.zeros(2)
numQPos = len(physicsSimulation.data.qpos)
numQVel = len(physicsSimulation.data.qvel)
sampleAgentsQPos = lambda: np.asarray(qPosInit) + np.random.uniform(
low=-qPosInitNoise, high=qPosInitNoise, size=numQPos)
sampleAgentsQVel = lambda: np.asarray(qVelInit) + np.random.uniform(
low=-qVelInitNoise, high=qVelInitNoise, size=numQVel)
minDistance = 0.2 + 2 * blockSize
isOverlap = IsOverlap(minDistance)
sampleBlockState = SampleBlockState(numBlocks, getBlockRandomPos,
getBlockSpeed, isOverlap)
reset = ResetUniformWithoutXPos(physicsSimulation, numAgents, numBlocks,
sampleAgentsQPos, sampleAgentsQVel,
sampleBlockState)
damping = 2.5
numSimulationFrames = int(0.1 / dt)
agentsMaxSpeedList = [wolfMaxSpeed] * numWolves + [sheepMaxSpeed
] * numSheeps
reshapeAction = ReshapeAction()
isTerminal = lambda state: False
transit = TransitionFunctionWithoutXPos(physicsSimulation, numAgents,
numSimulationFrames,
damping * dt * numSimulationFrames,
agentsMaxSpeedList,
visualizeMujoco, isTerminal,
reshapeAction)
maxRunningStepsToSample = 100
sampleTrajectory = SampleTrajectory(maxRunningStepsToSample, transit,
isTerminal, rewardFunc, reset)
observeOneAgent = lambda agentID: Observe(agentID, wolvesID, sheepsID,
blocksID, getPosFromAgentState,
getVelFromAgentState)
observe = lambda state: [
observeOneAgent(agentID)(state) for agentID in range(numAgents)
]
initObsForParams = observe(reset())
obsShape = [
initObsForParams[obsID].shape[0]
for obsID in range(len(initObsForParams))
]
worldDim = 2
actionDim = worldDim * 2 + 1
layerWidth = [128, 128]
# ------------ model ------------------------
buildMADDPGModels = BuildMADDPGModels(actionDim, numAgents, obsShape)
modelsList = [
buildMADDPGModels(layerWidth, agentID) for agentID in range(numAgents)
]
if hasWalls:
modelFolder = os.path.join(
dirName, '..', 'trainedModels', 'mujocoMADDPGeavluateWall',
'dt={}'.format(dt),
'hasWalls=' + str(hasWalls) + 'numBlocks=' + str(numBlocks) +
'numSheeps=' + str(numSheeps) + 'numWolves=' + str(numWolves) +
'timeconst=' + str(timeconst) + 'dampratio=' + str(dampratio) +
'individualRewardWolf=' + str(individualRewardWolf) +
'sheepSpeedMultiplier=' + str(sheepSpeedMultiplier))
individStr = 'individ' if individualRewardWolf else 'shared'
fileName = "maddpghasWalls={}{}wolves{}sheep{}blocks{}episodes{}stepSheepSpeed{}{}_agent".format(
hasWalls, numWolves, numSheeps, numBlocks, maxEpisode, maxTimeStep,
sheepSpeedMultiplier, individStr)
else:
modelFolder = os.path.join(
dirName, '..', 'trainedModels', 'mujocoMADDPG',
'numBlocks=' + str(numBlocks) + 'numSheeps=' + str(numSheeps) +
'numWolves=' + str(numWolves) + 'timeconst=' + str(timeconst) +
'dampratio=' + str(dampratio))
fileName = "maddpg{}wolves{}sheep{}blocks60000episodes25stepSheepSpeed1.0shared_agent".format(
numWolves, numSheeps, numBlocks)
modelPaths = [
os.path.join(modelFolder, fileName + str(i) + '60000eps')
for i in range(numAgents)
]
[
restoreVariables(model, path)
for model, path in zip(modelsList, modelPaths)
]
actOneStepOneModel = ActOneStep(actByPolicyTrainNoisy)
policy = lambda allAgentsStates: [
actOneStepOneModel(model, observe(allAgentsStates))
for model in modelsList
]
startTime = time.time()
trajList = []
for i in range(evalNum):
np.random.seed(i)
traj = sampleTrajectory(policy)
trajList.append(list(traj))
endTime = time.time()
print("Time taken {} seconds to generate {} trajectories".format(
(endTime - startTime), evalNum))
# saveTraj
if saveTraj:
trajectoryFolder = os.path.join(dirName, '..', 'trajectory',
'evluateWall')
if not os.path.exists(trajectoryFolder):
os.makedirs(trajectoryFolder)
trajectorySaveExtension = '.pickle'
fixedParameters = {'randomSeed': randomSeed, 'evalNum': evalNum}
generateTrajectorySavePath = GetSavePath(trajectoryFolder,
trajectorySaveExtension,
fixedParameters)
trajectorySavePath = generateTrajectorySavePath(parameterOneCondiiton)
saveToPickle(trajList, trajectorySavePath)
# visualize
if visualizeTraj:
entitiesColorList = [wolfColor] * numWolves + [
sheepColor
] * numSheeps + [blockColor] * numBlocks
render = Render(entitiesSizeList, entitiesColorList, numAgents,
getPosFromAgentState)
print(trajList[0][0], '!!!3', trajList[0][1])
trajToRender = np.concatenate(trajList)
render(trajToRender)
return endTime - startTime
def main():
stateDim = env.observation_space.shape[0]
actionDim = 7
buildModel = BuildModel(stateDim, actionDim)
layersWidths = [30]
writer, model = buildModel(layersWidths)
learningRate = 0.001
gamma = 0.99
trainModelBySASRQ = TrainModelBySASRQ(learningRate, gamma, writer)
paramUpdateInterval = 300
updateParameters = UpdateParameters(paramUpdateInterval)
model = resetTargetParamToTrainParam([model])[0]
trainModels = TrainDQNModel(getTargetQValue, trainModelBySASRQ, updateParameters, model)
epsilonMax = 0.9
epsilonIncrease = 0.0001
epsilonMin = 0
bufferSize = 10000
decayStartStep = bufferSize
getEpsilon = GetEpsilon(epsilonMax, epsilonMin, epsilonIncrease, decayStartStep)
actGreedyByModel = ActGreedyByModel(getTrainQValue, model)
actRandom = ActRandom(actionDim)
actByTrainNetEpsilonGreedy = ActByTrainNetEpsilonGreedy(getEpsilon, actGreedyByModel, actRandom)
minibatchSize = 128
learningStartBufferSize = minibatchSize
sampleFromMemory = SampleFromMemory(minibatchSize)
learnFromBuffer = LearnFromBuffer(learningStartBufferSize, sampleFromMemory, trainModels)
processAction = ProcessDiscretePendulumAction(actionDim)
transit = TransitGymPendulum(processAction)
getReward = RewardGymPendulum(angle_normalize, processAction)
sampleOneStep = SampleOneStep(transit, getReward)
runDQNTimeStep = RunTimeStep(actByTrainNetEpsilonGreedy, sampleOneStep, learnFromBuffer, observe)
reset = ResetGymPendulum(seed)
maxTimeStep = 200
runEpisode = RunEpisode(reset, runDQNTimeStep, maxTimeStep, isTerminalGymPendulum)
maxEpisode = 400
dqn = RunAlgorithm(runEpisode, maxEpisode)
replayBuffer = deque(maxlen=int(bufferSize))
meanRewardList, trajectory = dqn(replayBuffer)
trainedModel = trainModels.getTrainedModels()
# save Model
parameters = {'maxEpisode': maxEpisode, 'maxTimeStep': maxTimeStep, 'minibatchSize': minibatchSize, 'gamma': gamma,
'learningRate': learningRate, 'epsilonIncrease': epsilonIncrease , 'epsilonMin': epsilonMin}
modelSaveDirectory = "../trainedDQNModels"
modelSaveExtension = '.ckpt'
getSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension)
savePath = getSavePath(parameters)
with trainedModel.as_default():
saveVariables(trainedModel, savePath)
dirName = os.path.dirname(__file__)
trajectoryPath = os.path.join(dirName, '..', 'trajectory', 'pendulumDQNTrajectory.pickle')
saveToPickle(trajectory, trajectoryPath)
plotResult = True
if plotResult:
plt.plot(list(range(maxEpisode)), meanRewardList)
plt.show()
showDemo = False
if showDemo:
visualize = VisualizeGymPendulum()
visualize(trajectory)
def main():
env = MtCarDiscreteEnvSetup()
visualize = visualizeMtCarDiscrete()
reset = resetMtCarDiscrete(1234)
transition = MtCarDiscreteTransition()
rewardMtCar = MtCarDiscreteReward()
isterminal = MtCarDiscreteIsTerminal()
statesdim = env.observation_space.shape[0]
actiondim = env.action_space.n
replaybuffer = deque(maxlen=buffersize)
runepsilon = initepsilon
totalrewards = []
meanreward = []
trajectory = []
totalreward = []
buildmodel = BuildModel(statesdim, actiondim)
Writer, DQNmodel = buildmodel(numberlayers)
replaceParameters = ReplaceParameters(replaceiter)
trainModel = TrainModel(learningRate, gamma, Writer)
trainDQNmodel = TrainDQNmodel(replaceParameters, trainModel, DQNmodel)
learn = Learn(buffersize, batchsize, trainDQNmodel, actiondim)
for episode in range(EPISODE):
state = reset()
rewards = 0
runtime = 0
while True:
action = learn.Getaction(DQNmodel, runepsilon, state)
nextstate = transition(state, action)
done = isterminal(nextstate)
reward = rewardMtCar(state, action, nextstate, done)
learn.ReplayMemory(replaybuffer, state, action, reward, nextstate,
done)
trajectory.append((state, action, reward, nextstate))
rewards += reward
state = nextstate
runtime += 1
if runtime == 200:
totalrewards.append(rewards)
totalreward.append(rewards)
runtime = 0
print('episode: ', episode, 'reward:', rewards, 'epsilon:',
runepsilon)
break
if done:
totalrewards.append(rewards)
totalreward.append(rewards)
print('episode: ', episode, 'reward:', rewards, 'epsilon:',
runepsilon)
break
runepsilon = epsilonDec(runepsilon, minepsilon, epsilondec)
if episode % 100 == 0:
meanreward.append(np.mean(totalreward))
print('episode: ', episode, 'meanreward:', np.mean(totalreward))
totalreward = []
episode = 100 * (np.arange(len(meanreward)))
plt.plot(episode, meanreward)
plt.xlabel('episode')
plt.ylabel('rewards')
plt.ylim([-200, -50])
plt.show()
# save Model
modelIndex = 0
DQNFixedParam = {'DQNmodel': modelIndex}
parameters = {
'env': ENV_NAME,
'Eps': EPISODE,
'batch': batchsize,
'buffersize': buffersize,
'gam': gamma,
'learningRate': learningRate,
'replaceiter': replaceiter,
'epsilondec': epsilondec,
'minepsilon': minepsilon,
'initepsilon': initepsilon
}
modelSaveDirectory = "/path/to/logs/trainedDQNModels"
modelSaveExtension = '.ckpt'
getSavePath = GetSavePath(modelSaveDirectory, modelSaveExtension,
DQNFixedParam)
savePathDQN = getSavePath(parameters)
with DQNmodel.as_default():
saveVariables(DQNmodel, savePathDQN)
dirName = os.path.dirname(__file__)
trajectoryPath = os.path.join(dirName, 'trajectory',
'mountCarTrajectory.pickle')
saveToPickle(trajectory, trajectoryPath)
