environment = CartPoleEnvironment()
experiment = Experiment(environment, agent)
# cut off last two state dimensions
indexer = IndexingAdapter([0, 1], None)
experiment.addAdapter(indexer)
# add normalization adapter
normalizer = NormalizingAdapter(scaleActions=[(-50, 50)])
experiment.addAdapter(normalizer)
# # add e-greedy exploration
# explorer = EpsilonGreedyExplorer(0.4, episodeCount=500)
# experiment.addAdapter(explorer)
experiment.runEpisodes(10)
agent.forget()
# explorer.decay = 0.999
# renderer = CartPoleRenderer()
# environment.renderer = renderer
# renderer.start()
# run experiment
for i in range(100):
experiment.runEpisodes(5)
agent.learn()
# agent.forget()
valdata = experiment.evaluateEpisodes(10, visualize=True)
print "mean return", mean_return
raise SystemExit
# add e-greedy exploration
explorer = EpsilonGreedyExplorer(0.3, episodeCount=1000)
experiment.addAdapter(explorer)
if os.path.exists('cart_play.saved'):
os.remove('cart_play.saved')
# run experiment
for i in range(1000):
valdata = experiment.evaluateEpisodes(20, visualize=True)
mean_return = mean([sum(v.rewards) for v in valdata])
experiment.runEpisodes(1)
agent.learn()
# agent.history.truncate(20)
# agent.forget()
# save file after each learning step
f = open('cart_play.saved', 'w')
cPickle.dump(agent, f)
f.close()
print normalizer.minStates, normalizer.maxStates
print "params", agent.estimator.models[0].W
print "exploration", explorer.epsilon
print "mean return", mean_return
print "num episodes", len(agent.history)
print "num total samples", agent.history.numTotalSamples()
import numpy as np # create agent, environment, renderer, experiment agent = APIAgent(faClass=LWPRFA, resetFA=True, vectorblock=False) agent.gamma = 2. agent.alpha = 1.0 agent.iterations = 1 agent.presentations = 1 environment = MirrorEnvironment() experiment = Experiment(environment, agent) # add bas adapter bas = BinaryActionSearchAdapter(3., 4., 10) experiment.addAdapter(bas) # add e-greedy exploration # explorer = EpsilonGreedyExplorer(0.5, episodeCount=10000) # experiment.addAdapter(explorer) # run experiment valdata = experiment.evaluateEpisodes(1000) print "mean rewards:", np.mean([sum(e.rewards) for e in valdata]) #, "exploration:", explorer.epsilon # print "exploration:", explorer.epsilon experiment.runEpisodes(10000) agent.learn() valdata = experiment.evaluateEpisodes(1000) print "mean rewards:", np.mean([sum(e.rewards) for e in valdata]) #, "exploration:", explorer.epsilon
from dopamine.environments import CartPoleEnvironment, CartPoleRenderer
from dopamine.agents import FiniteDifferenceAgent, NNController
from dopamine.adapters import IndexingAdapter, NormalizingAdapter
from dopamine.experiments import Experiment
from numpy import *
environment = CartPoleEnvironment()
environment.centerCart = False
agent = FiniteDifferenceAgent(controllerClass=NNController)
experiment = Experiment(environment, agent)
# cut off last two state dimensions
indexer = IndexingAdapter([0, 1], None)
experiment.addAdapter(indexer)
# add normalization adapter
normalizer = NormalizingAdapter(scaleActions=[(-50, 50)])
experiment.addAdapter(normalizer)
# run experiment
for i in range(1000):
experiment.runEpisodes(50)
agent.learn()
agent.forget()
valdata = experiment.evaluateEpisodes(10)
print "mean return", mean([sum(v.rewards) for v in valdata])