agent.learner.gd.alpha = 0.3 #step size of \mu adaption
agent.learner.gdSig.alpha = 0.15 #step size of \sigma adaption
agent.learner.gd.momentum = 0.0
batch=2 #number of samples per gradient estimate (was: 2; more here due to stochastic setting)
#create experiment
experiment = EpisodicExperiment(task, agent)
prnts=1 #frequency of console output
epis=2000/batch/prnts
#actual roll outs
filename="dataSPLA08NoRew"+repr(int(random.random()*1000000.0))+".dat"
wf = open(filename, 'wb')
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch) #execute #batch episodes
agent.learn() #learn from the gather experience
agent.reset() #reset agent and environment
#print out related data
stp = (updates+1)*batch*prnts
print "Step: ", runs, "/", stp, "Best: ", agent.learner.best, "Base: ", agent.learner.baseline, "Reward: ", agent.learner.reward
wf.write(repr(stp)+"\n")
wf.write(repr(agent.learner.baseline[0])+"\n")
if useGraphics:
pl.addData(0,float(stp),agent.learner.baseline)
pl.addData(1,float(stp),agent.learner.best)
pl.update()
#if updates/100 == float(updates)/100.0:
# saveWeights("walk.wgt", agent.learner.original)
wf.close()
agent.learner.gdSig.alpha = 0.1
agent.learner.gd.momentum = 0.9
agent.learner.epsilon = 6.0
agent.learner.initSigmas()
# agent.learner.rprop = True
experiment = EpisodicExperiment(task, agent)
batch = 16
prnts = 10
epis = 50000 / batch / prnts
save = False
rl = []
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch)
agent.learn()
agent.reset()
print "Parameters: ", agent.learner.original
print "Epsilon : ", agent.learner.sigList
print "Step: ", runs, "/", (
updates + 1
) * batch * prnts, "Best: ", agent.learner.best, "Base: ", agent.learner.baseline, "Reward: ", agent.learner.reward
print ""
rl.append(float(agent.learner.baseline))
if save:
fnStart = "dataCartMom"
fnExp = repr(int(agent.learner.gd.alpha * 100)) + "m" + repr(
int(agent.learner.gdSig.alpha * 100)) + "s" + repr(
batch / 2) + "b" + repr(int(agent.learner.epsilon * 10)) + "e"
fnIdent = "SPLA" + repr(int(random.random() * 1000000.0))
filename = fnStart + fnExp + fnIdent + ".dat"
agent.learner.gdSig.alpha = 0.085 #step size of \sigma adaption
agent.learner.gd.momentum = 0.0
#Loading weights
if loadNet:
agent.learner.original = loadWeights("grasp.wgt")
agent.learner.gd.init(agent.learner.original)
agent.learner.epsilon = 0.2
agent.learner.initSigmas()
batch = 2 #number of samples per gradient estimate
#create experiment
experiment = EpisodicExperiment(task, agent)
prnts = 1 #frequency of console output
epis = 5000000 / batch / prnts
#actual roll outs
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch) #execute batch episodes
agent.learn() #learn from the gather experience
agent.reset() #reset agent and environment
#print out related data
print "Step: ", runs, "/", (
updates + 1) * batch * prnts, "Best: ", agent.learner.best,
print "Base: ", agent.learner.baseline, "Reward: ", agent.learner.reward
#Saving weights
if saveNet:
if updates / 100 == float(updates) / 100.0:
saveWeights(saveName, agent.learner.original)
agent.learner.gdSig.alpha = 0.1
agent.learner.gd.momentum = 0.9
agent.learner.epsilon = 6.0
agent.learner.initSigmas()
# agent.learner.rprop = True
experiment = EpisodicExperiment(task, agent)
batch = 16
prnts = 10
epis = 50000 / batch / prnts
save = False
rl = []
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch)
agent.learn()
agent.reset()
print "Parameters: ", agent.learner.original
print "Epsilon : ", agent.learner.sigList
print "Step: ", runs, "/", (
updates + 1
) * batch * prnts, "Best: ", agent.learner.best, "Base: ", agent.learner.baseline, "Reward: ", agent.learner.reward
print ""
rl.append(float(agent.learner.baseline))
if save:
fnStart = "dataCartMom"
fnExp = (
repr(int(agent.learner.gd.alpha * 100))
+ "m"
+ repr(int(agent.learner.gdSig.alpha * 100))
+ "s"