def cycleTrain(saveDir, genPath, saveAgentName):
while True:
preprocDf = preprocData(df)
createGenerator(preprocDf,
featureList=hkFeatList,
saveDir=saveDir,
saveName=genName)
trainDf = preprocDf.tail(1232).head(1032)
backTestDf = preprocDf.tail(1432).head(232)
lastSaveEp = trainAgent(trainDf=trainDf,
backTestDf=backTestDf,
genPath=genPath,
saveAgentDir=saveDir,
saveAgentName=saveAgentName)
#####
# break
#####
sumReward = evaluateAgent(backTestDf=backTestDf,
genPath=genPath,
agentDir=saveDir,
agentName=saveAgentName)
if lastSaveEp > 0 and sumReward > 0:
break
else:
print("Backtest failed. Retrain.")
reset_keras()
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,
stoplossPuncts=60, takeprofitPuncts=120, renderFlag=True)
# 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()
def useAgent(preprocDf, hkFeatList, saveDir, genPath, agentName,
timeConstraint):
###############################
# use agent on test
###############################
useCpu(nThreads=8, nCores=8)
testDf = preprocDf.tail(232)
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 = CompositeEnv(testDf,
openerPriceDiffGenerator,
buyerPriceDiffGenerator,
sellerPriceDiffGenerator,
startDeposit=300,
lotSize=0.1,
lotCoef=100000,
spread=18,
spreadCoef=0.00001,
renderFlag=True,
renderDir=saveDir,
renderName="testDealsPlot")
# 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()
dealsStatistics = agent.use_agent(testEnv, timeConstraint=timeConstraint)
endTime = datetime.now()
print("Use time: {}".format(endTime - startTime))
reset_keras()
sumRew = 0
cumulativeReward = []
for i in range(len(dealsStatistics)):
sumRew += dealsStatistics[i]
cumulativeReward.append(sumRew)
plt.plot([x for x in range(len(cumulativeReward))], cumulativeReward)
plt.show()
def cycleTrain(symbol, timeframe, dataUpdater, saveDir, genPath,
saveAgentName):
while True:
dataUpdater.partialUpdate(terminal, symbol, timeframe)
df = SymbolDataManager().getData(symbol, timeframe)
preprocDf = preprocData(df)
createGenerator(preprocDf,
featureList=hkFeatList,
saveDir=saveDir,
saveName=genName)
trainDf = preprocDf.tail(2032).tail(1032)
backTestDf = preprocDf.tail(2032).head(1032)
lastSaveEp = trainAgent(trainDf=trainDf,
backTestDf=backTestDf,
genPath=genPath,
saveAgentDir=saveDir,
saveAgentName=saveAgentName)
##############
reset_keras()
##############
#####
# break
#####
sumReward = evaluateAgent(backTestDf=backTestDf,
genPath=genPath,
agentDir=saveDir,
agentName=saveAgentName)
if lastSaveEp > 0 and sumReward > -110:
print("Backtest's succesfull. Stop Training.")
break
else:
print("Backtest failed. Retrain.")
#############
#weighted train/back reward
"""backReward = evaluateAgent(backTestDf=backTestDf, genPath=genPath, agentDir=saveDir, agentName=saveAgentName)
trainReward = evaluateAgent(backTestDf=trainDf, genPath=genPath, agentDir=saveDir, agentName=saveAgentName)
backW = 0.8 #1.0 - (len(backTestDf) / (len(backTestDf) + len(trainDf)))
trainW = 0.2 #1.0 - (len(trainDf) / (len(backTestDf) + len(trainDf)))
weightedReward = (backW * backReward + trainW * trainReward) / 2
if lastSaveEp > 0 and weightedReward > 0:
print("Backtest's succesfull. Stop Training.")
break
else:
print("Backtest failed. Retrain.")"""
#############
reset_keras()
def collectStatistics(symbol, nExperiment, trainDf, backTestDf, testDf, startDeposit=300, lotSize=0.1, lotCoef=100000, spread=18, spreadCoef=0.00001):
reset_keras()
openerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=priceFeatList).loadGenerator("./MSDiffGen.pkl")
buyerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=priceFeatList).loadGenerator("./MSDiffGen.pkl")
sellerPriceDiffGenerator = MultiScalerDiffGenerator(featureList=priceFeatList).loadGenerator("./MSDiffGen.pkl")
trainEnv = CompositeEnv(trainDf, openerPriceDiffGenerator, buyerPriceDiffGenerator, sellerPriceDiffGenerator,
startDeposit=startDeposit, lotSize=lotSize, lotCoef=lotCoef, spread=spread, spreadCoef=spreadCoef, renderFlag=True)
backTestEnv = CompositeEnv(backTestDf, openerPriceDiffGenerator, buyerPriceDiffGenerator, sellerPriceDiffGenerator,
startDeposit=startDeposit, lotSize=lotSize, lotCoef=lotCoef, spread=spread, spreadCoef=spreadCoef, renderFlag=True)
forwardTestEnv = CompositeEnv(testDf, openerPriceDiffGenerator, buyerPriceDiffGenerator, sellerPriceDiffGenerator,
startDeposit=startDeposit, lotSize=lotSize, lotCoef=lotCoef, spread=spread, spreadCoef=spreadCoef, renderFlag=True)
# get size of state and action from environment
openerAgent = DQNAgent(trainEnv.observation_space["opener"], trainEnv.action_space["opener"].n,
memorySize=2500, batch_size=100, train_start=200, epsilon_min=0.05, epsilon=1,
epsilon_decay=0.9995, learning_rate=0.0005)
buyerAgent = DQNAgent(trainEnv.observation_space["buyer"], trainEnv.action_space["buyer"].n,
memorySize=5000, batch_size=200, train_start=300, epsilon_min=0.05, epsilon=1,
epsilon_decay=1.0, learning_rate=0.0005)
sellerAgent = DQNAgent(trainEnv.observation_space["seller"], trainEnv.action_space["seller"].n,
memorySize=5000, batch_size=200, train_start=300, epsilon_min=0.05, epsilon=1,
epsilon_decay=1.0, learning_rate=0.0005)
agent = CompositeAgent(openerAgent, buyerAgent, sellerAgent)
continueFit = False
trainDealsStatistics = []
testDealsStatistics = []
backDealsStatistics = []
for i in range(20):
agent.fit_agent(env=trainEnv, backTestEnv=None, nEpisodes=1, nWarmUp=0,
uniformEps=False, synEps=True, plotScores=True, saveBest=False, saveFreq=111, continueFit=continueFit)
continueFit = True
trainDealsStatistics.append( agent.use_agent(trainEnv) )
testDealsStatistics.append( agent.use_agent(forwardTestEnv) )
backDealsStatistics.append( agent.use_agent(backTestEnv) )
with open("./" + "trainDealsStatistics_{}_{}.pkl".format(symbol, nExperiment), mode="wb") as dealsFile:
joblib.dump(trainDealsStatistics, dealsFile)
with open("./" + "testDealsStatistics_{}_{}.pkl".format(symbol, nExperiment), mode="wb") as dealsFile:
joblib.dump(testDealsStatistics, dealsFile)
with open("./" + "backDealsStatistics_{}_{}.pkl".format(symbol, nExperiment), mode="wb") as dealsFile:
joblib.dump(backDealsStatistics, dealsFile)
pass
nWarmUp=0,
uniformEps=False,
synEps=False,
plotScores=False,
saveBest=True,
saveFreq=1)
endTime = datetime.now()
print("Training finished. Total time: {}".format(endTime - startTime))
###############################
# use agent
###############################
from avera.utils.keras_utils import reset_keras
reset_keras()
#testDf = modDf.tail(12000).head(2000)
#testDf = modDf.tail(10000).head(8000)
testDf = modDf.tail(35000).head(5000)
#testDf = modDf.tail(110000).head(10000) #back
openerPriceDiffGenerator = MultiScalerDiffGenerator(
featureList=hkFeatList).loadGenerator("./MSDiffGen.pkl")
buyerPriceDiffGenerator = MultiScalerDiffGenerator(
featureList=hkFeatList).loadGenerator("./MSDiffGen.pkl")
sellerPriceDiffGenerator = MultiScalerDiffGenerator(
featureList=hkFeatList).loadGenerator("./MSDiffGen.pkl")
#openerPriceDiffGenerator.setFitMode(False)
#buyerPriceDiffGenerator.setFitMode(False)