# 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 "exploration", explorer.epsilon
print "mean return", mean([sum(v.rewards) for v in valdata])
print "num episodes", len(agent.history)
# print "num total samples", agent.history.numTotalSamples()
# force setup here already to initiate pretraining
experiment.setup()
# environment.renderer = renderer
# renderer.start()
# experiment.runEpisodes(4)
# run experiment
for i in range(5000):
experiment.runEpisodes(10)
agent.learn()
agent.history.keepBest(50)
# agent.forget()
valdata = experiment.evaluateEpisodes(10, visualize=False)
if renderer.isAlive():
environment.renderer = renderer
experiment.evaluateEpisodes(1, visualize=False)
environment.renderer = None
print i
print "mean return", mean([sum(v.rewards) for v in valdata])
if mean([sum(v.rewards) for v in valdata]) > 1.75*maxSteps:
if not renderer.isAlive():
renderer.start()
pass
print "avg. episode length", mean([len(v) for v in valdata])
print "exploration variance", explorer.sigma
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
# cut off last two state dimensions
indexer = IndexingAdapter([0, 1], None)
experiment.addAdapter(indexer)
# add normalization adapter
normalizer = NormalizingAdapter()
experiment.addAdapter(normalizer)
# add e-greedy exploration
explorer = EpsilonGreedyExplorer(0.2, episodeCount=100)
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(1)
agent.learn()
valdata = experiment.evaluateEpisodes(5)
print "exploration", explorer.epsilon
print "mean return", mean([sum(v.rewards) for v in valdata])
print "num episodes", len(agent.history)