obs[-2 + self.extraRandoms] = 0.1 * sin(angle1) + cartpos
if self.numPoles == 2:
if self.markov:
angle2 = obs[3]
else:
angle2 = obs[1]
obs[-3 + self.extraRandoms] = 0.05 * cos(angle2) + cartpos
obs[-4 + self.extraRandoms] = 0.05 * sin(angle2) + cartpos
if self.extraRandoms > 0:
obs[-self.extraRandoms:] = randn(self.extraRandoms)
if self.verbose:
print('obs', obs)
return obs
def performAction(self, action):
if self.verbose:
print('act', action)
impl.performAction(action[0])
self.addReward()
if __name__ == '__main__':
from pybrain.rl import EpisodicExperiment
from pybrain.rl.agents import FlatNetworkAgent
x = FastCartPoleTask()
a = FlatNetworkAgent(x.outdim, x.indim)
e = EpisodicExperiment(x, a)
e.doEpisodes(2)
#net.initParams(0.0)
# create agent
agent = StateDependentAgent(net, ENAC())
agent.learner.gd.rprop = True
# only relevant for RP
agent.learner.gd.deltamin = 0.0001
#agent.learner.gd.deltanull = 0.05
# only relevant for BP
agent.learner.gd.alpha = 0.01
agent.learner.gd.momentum = 0.9
agent.actaspg = False
# create experiment
experiment = EpisodicExperiment(task, agent)
# print weights at beginning
print agent.module.params
rewards = []
if useGraphics:
figure()
ion()
pl = MultilinePlotter(autoscale=1.2, xlim=[0, 50], ylim=[0, 1])
pl.setLineStyle(linewidth=2)
# queued version
# experiment._fillQueue(30)
# while True:
# experiment._stepQueueLoop()
if self.numPoles == 2:
if self.markov:
angle2 = obs[3]
else:
angle2 = obs[1]
obs[-3 + self.extraRandoms] = 0.05 * cos(angle2) + cartpos
obs[-4 + self.extraRandoms] = 0.05 * sin(angle2) + cartpos
if self.extraRandoms > 0:
obs[-self.extraRandoms :] = randn(self.extraRandoms)
if self.verbose:
print "obs", obs
return obs
def performAction(self, action):
if self.verbose:
print "act", action
impl.performAction(action[0])
self.addReward()
if __name__ == "__main__":
from pybrain.rl import EpisodicExperiment
from pybrain.rl.agents import FlatNetworkAgent
x = FastCartPoleTask()
a = FlatNetworkAgent(x.outdim, x.indim)
e = EpisodicExperiment(x, a)
e.doEpisodes(2)
# create agent
agent = StateDependentAgent(net, ENAC())
agent.learner.gd.rprop = True
# only relevant for RP
agent.learner.gd.deltamin = 0.0001
#agent.learner.gd.deltanull = 0.05
# only relevant for BP
agent.learner.gd.alpha = 0.01
agent.learner.gd.momentum = 0.9
agent.actaspg = False
# create experiment
experiment = EpisodicExperiment(task, agent)
# print weights at beginning
print agent.module.params
rewards = []
if useGraphics:
figure()
ion()
pl = MultilinePlotter(autoscale=1.2, xlim=[0, 50], ylim=[0, 1])
pl.setLineStyle(linewidth=2)
# queued version
# experiment._fillQueue(30)
# while True:
# experiment._stepQueueLoop()
