def __init__(self):
self.this_ai_count = planet_environment.ai_count
self.finished = False
planet_environment.ai_count += 1
self.nextMove = (0, 0, 0)
self.state = None
network = buildNetwork(INPUT_NEURON_COUNT, 100, 3)
enac_learner = ENAC()
learning_agent = LearningAgent(network, enac_learner)
self.experiment = planet_experiment(episodic_planet_task(self),
learning_agent)
self.experiment.task.clipping = False
def mlDriver(cv, stateTransfer, actionTransfer):
#parameter setup
#dimensionality of state argument (could be less than stateTransfer)
stateDim = 352
#Number of moves possible
numMoves = 361
env = SettleEnv(cv, stateTransfer, actionTransfer)
task = SettleTask(env)
controller = RestrictedActionValueNetwork(stateDim, numMoves, env)
learner = NFQ()
learner.explorer = EpsilonHackedExplorer(env)
agent = LearningAgent(controller, learner)
experiment = EpisodicExperiment(task, agent)
while True:
experiment.doEpisodes(10)
print "Done with experiments"
agent.learn()
print "Learned"
agent.reset()
print "Cycled"
sys.exit('please give 4 parameters. run: "python play_catpole.py <p1> <p2> <p3> <p4>"\n')
# create environment
env = CartPoleEnvironment()
env.setRenderer(CartPoleRenderer())
env.getRenderer().start()
env.delay = (episodes == 1)
# create task
task = BalanceTask(env, epilen)
# create controller network
net = buildNetwork(4, 1, bias=False)
# create agent and set parameters from command line
agent = LearningAgent(net, None)
agent.module._setParameters([float(sys.argv[1]), float(sys.argv[2]), float(sys.argv[3]), float(sys.argv[4])])
# create experiment
experiment = EpisodicExperiment(task, agent)
experiment.doEpisodes(episodes)
# run environment
ret = []
for n in range(agent.history.getNumSequences()):
returns = agent.history.getSequence(n)
reward = returns[2]
ret.append( sum(reward, 0).item() )
# print results
print ret, "mean:",mean(ret)
def __init__(self, n_input,actions, alpha=0.5, gamma=0.99, qlambda=0.9,explorer=EpsilonGreedyExplorer(epsilon=0.20,decay=1)):
CompleteLearner.__init__(self,actions)
controller = ActionValueNetwork(dimState=n_input, numActions=len(actions))
learner = NFQ()
learner.explorer = explorer
self.learning_agent = LearningAgent(controller, learner)
