def useAgent(symbol, timeframe, terminal, dataUpdater, dataManager,
hkFeatList, saveDir, genPath, agentName, timeConstraint):
###############################
# use agent on test
###############################
useCpu(nThreads=8, nCores=8)
openerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=hkFeatList).loadGenerator(genPath)
buyerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=hkFeatList).loadGenerator(genPath)
sellerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=hkFeatList).loadGenerator(genPath)
openerPriceDiffGenerator.setFitMode(False)
buyerPriceDiffGenerator.setFitMode(False)
sellerPriceDiffGenerator.setFitMode(False)
testEnv = RealCompositeEnv(symbol, timeframe, terminal, dataUpdater, dataManager,
openerPriceDiffGenerator, buyerPriceDiffGenerator, sellerPriceDiffGenerator,
startDeposit=300, lotSize=0.1, lotCoef=100000, stopType="adaptive", takeType="adaptive",
stoplossPuncts=100, takeprofitPuncts=200, stopPos=2, takePos=1, maxLoss=40000,
maxTake=40000, riskPoints=110, riskLevels=5, renderFlag=False)
# get size of state and action from environment
openerAgent = DQNAgent(testEnv.observation_space["opener"], testEnv.action_space["opener"].n)
buyerAgent = DQNAgent(testEnv.observation_space["buyer"], testEnv.action_space["buyer"].n)
sellerAgent = DQNAgent(testEnv.observation_space["seller"], testEnv.action_space["seller"].n)
agent = CompositeAgent(openerAgent, buyerAgent, sellerAgent)
agent = agent.load_agent(saveDir, agentName, dropSupportModel=True)
# agent = agent.load_agent("./", "checkpoint_composite")
print("start using agent")
startTime = datetime.now()
agent.use_agent(testEnv, timeConstraint=timeConstraint)
endTime = datetime.now()
print("Use time: {}".format(endTime - startTime))
reset_keras()
trainDf = modDf.tail(30000).head(27000)
trainEnv = CompositeEnv(trainDf, openerCompositeGenerator, buyerCompositeGenerator, sellerCompositeGenerator,
startDeposit=300, lotSize=0.1, lotCoef=100000, spread=18, spreadCoef=0.00001,
stopType="const", takeType="const", stopPos=2, takePos=1, maxLoss=20000, maxTake=20000,
stoplossPuncts=20000, takeprofitPuncts=20000, riskPoints=110, riskLevels=5, parallelOpener=False,
renderFlag=True, renderDir="../models/", renderName="train_plot")
backTestDf = modDf.tail(33000).head(3000)
#backTestDf = modDf.head(50000).tail(3192).head(1192)
backTestEnv = CompositeEnv(backTestDf, openerCompositeGenerator, buyerCompositeGenerator, sellerCompositeGenerator,
startDeposit=300, lotSize=0.1, lotCoef=100000, spread=18, spreadCoef=0.00001,
stopType="adaptive", takeType="adaptive", stopPos=1, takePos=1, maxLoss=20000, maxTake=20000,
stoplossPuncts=20000, takeprofitPuncts=20000, riskPoints=110, riskLevels=5, parallelOpener=False,
renderDir="../models/", renderName="back_plot")
# get size of state and action from environment
openerAgent = DQNAgent(trainEnv.observation_space["opener"], trainEnv.action_space["opener"].n,
memorySize=40000, batch_size=20, train_start=20000, epsilon_min=0.2, epsilon=1, discount_factor=0.99,
epsilon_decay=0.9999, learning_rate=0.0001)
buyerAgent = DQNAgent(trainEnv.observation_space["buyer"], trainEnv.action_space["buyer"].n,
memorySize=40000, batch_size=20, train_start=10000, epsilon_min=0.2, epsilon=1, discount_factor=0.99,
epsilon_decay=0.9999, learning_rate=0.0001)
sellerAgent = DQNAgent(trainEnv.observation_space["seller"], trainEnv.action_space["seller"].n,
memorySize=40000, batch_size=20, train_start=10000, epsilon_min=0.2, epsilon=1, discount_factor=0.99,
epsilon_decay=0.9999, learning_rate=0.0001)
agent = CompositeAgent(openerAgent, buyerAgent, sellerAgent)
###################################
#agent = agent.load_agent("../models/", "best_composite")
#agent.agents["opener"].epsilon_min = 0.1
#agent.agents["buy"].epsilon_min = 0.1
#agent.agents["sell"].epsilon_min = 0.1
#agent = agent.load_agent("../models/", "checkpoint_composite")
takeType="adaptive",
stopPos=1,
takePos=1,
maxLoss=20000,
maxTake=20000,
stoplossPuncts=2000,
takeprofitPuncts=2000,
riskPoints=110,
riskLevels=5,
parallelOpener=False,
renderDir="../models/",
renderName="test_plot")
# get size of state and action from environment
openerAgent = DQNAgent(testEnv.observation_space["opener"],
testEnv.action_space["opener"].n,
fillMemoryByPretrainedModel=True)
buyerAgent = DQNAgent(testEnv.observation_space["buyer"],
testEnv.action_space["buyer"].n,
fillMemoryByPretrainedModel=True)
sellerAgent = DQNAgent(testEnv.observation_space["seller"],
testEnv.action_space["seller"].n,
fillMemoryByPretrainedModel=True)
#buyerAgent = AlwaysHoldAgent(testEnv.observation_space["opener"], testEnv.action_space["opener"].n, agentName="buyer")
#sellerAgent = AlwaysHoldAgent(testEnv.observation_space["seller"], testEnv.action_space["seller"].n, agentName="seller")
agent = CompositeAgent(openerAgent, buyerAgent, sellerAgent)
#agent = agent.load_agent("../models/", "best_composite", dropSupportModel=True)
#agent = agent.loadPretrainedAgents(dir="../models/", baseName="super_resnet_34_{}_{}".format(symbol, timeframe))
agent = agent.loadPretrainedAgents(dir="../models/", baseName="qrl_resnet_34")
#agent = agent.load_agent("../models/", "checkpoint_composite")
print("start using agent")