def createAgentCont(dimAct,dimState,numBatchToKeep, numIterPerTrain):
from pybrain.rl.learners import NFQ
sizeBatch = numIterPerTrain
learner = NFQ(sizeBatch, numBatchToKeep) # Use neuro-fitted Q learning (use Q learning with neural network instead of lookup table)
# Create a neural network model with dimState inputs and dimAct outputs
# Then network itself has dimState + dimAct input and 1 output
numHidden = 20
print('Using this many hidden layer neurons: ', numHidden)
moduleNet = ActionValueNetwork(dimState, dimAct, numHidden); moduleNet.name = 'moduleNet'
# Create a learning agent, using both the module and the learner
agent = LearningAgent(moduleNet, learner)
return agent
def initExperiment(learnalg='Q',
history=None,
binEdges='10s',
scriptfile='./rlRunExperiment_v2.pl',
resetscript='./rlResetExperiment.pl'):
if binEdges == '10s':
centerBinEdges = centerBinEdges_10s
elif binEdges == '30s':
centerBinEdges = centerBinEdges_30s
elif binEdges == 'lessperturbed':
centerBinEdges = centerBinEdges_10s_lessperturbed
elif binEdges is None:
centerBinEdges = None
else:
raise Exception("No bins for given binEdges setting")
env = OmnetEnvironment(centerBinEdges, scriptfile, resetscript)
if history is not None:
env.data = history['data']
task = OmnetTask(env, centerBinEdges)
if history is not None:
task.allrewards = history['rewards']
if learnalg == 'Q':
nstates = env.numSensorBins**env.numSensors
if history is None:
av_table = ActionValueTable(nstates, env.numActions)
av_table.initialize(1.)
else:
av_table = history['av_table']
learner = Q(0.1, 0.9) # alpha, gamma
learner._setExplorer(EpsilonGreedyExplorer(0.05)) # epsilon
elif learnalg == 'NFQ':
av_table = ActionValueNetwork(env.numSensors, env.numActions)
learner = NFQ()
else:
raise Exception("learnalg unknown")
agent = LearningAgent(av_table, learner)
experiment = Experiment(task, agent)
if history is None:
experiment.nruns = 0
else:
experiment.nruns = history['nruns']
return experiment
def main():
# 2048の全ての状態を保存するのは無理でしょ.
# 14^16通りの状態があるよね.
#controller = ActionValueTable(16, 4)
#learner = Q()
#controller.initialize(1.)
controller = ActionValueNetwork(16, 4)
learner = NFQ()
#learner._setExplorer(EpsilonGreedyExplorer(0.0))
agent = LearningAgent(controller, learner)
score_list = []
for i in range(10000):
# if os.path.exists('./agent.dump'):
# with open('./agent.dump') as f:
# agent = pickle.load(f)
print i, 'playing ...'
score = play(agent)
score_list.append(score)
# ここで,
# TypeError: only length-1 arrays can be converted to Python scalars
# pybrain/rl/learners/valuebased/q.py
# => learnerをQからNFQにしたら行けた.
# => http://stackoverflow.com/questions/23755927/pybrain-training-a-actionvaluenetwork-doesnt-properly-work
print i, 'learning ...'
agent.learn()
agent.reset()
print i, 'evaluate sample ...'
data = [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 2], [0, 0, 0, 2]]
agent.integrateObservation(numpy.array(data).ravel())
move = agent.getAction()
print " ", i, int(
numpy.mean(score_list)), max(score_list), move
if i % 20 == 0:
print i, 'saving ...'
with open('./agent.dump', 'w') as f:
pickle.dump(agent, f)
with open('./score.dump', 'w') as f:
pickle.dump(score_list, f)
from scipy import *
import sys, time
from pybrain.rl.learners.valuebased import ActionValueNetwork
from pybrain.rl.agents import LearningAgent
from pybrain.rl.learners import Q, SARSA, NFQ
from pybrain.rl.experiments.episodic import EpisodicExperiment
from pybrain.rl.environments import Task
from tasktest import TestTask
from envtest import TestEnv
env = TestEnv()
task = TestTask(env)
controller = ActionValueNetwork(200, 3)
learner = NFQ()
agent = LearningAgent(controller, learner)
experiment = EpisodicExperiment(task, agent)
i = 0
while True:
experiment.doEpisodes(10)
print "Learning"
agent.learn()
agent.reset()
i += 1
print "Cycle: %d" % i
if i > 60:
agent.learning = False
from pybrain.rl.environments.cartpole import CartPoleEnvironment, DiscreteBalanceTask
from pybrain.rl.agents import LearningAgent
from pybrain.rl.learners.valuebased import NFQ, ActionValueNetwork
from pybrain.rl.experiments import EpisodicExperiment
from training import NFQTraining
task = DiscreteBalanceTask(CartPoleEnvironment(), 100)
action_value_function = ActionValueNetwork(4, 3,
name='CartPoleNFQActionValueNetwork')
learner = NFQ()
#learner.gamma = 0.99
learner.explorer.epsilon = 0.4
task.discount = learner.gamma
agent = LearningAgent(action_value_function, learner)
performance_agent = LearningAgent(action_value_function, None)
experiment = EpisodicExperiment(task, agent)
tr = NFQTraining('cartpole_nfq', experiment, performance_agent)
tr.train(7000, performance_interval=1, n_performance_episodes=5)
from numpy import array, arange, meshgrid, pi, zeros, mean
from matplotlib import pyplot as plt
# switch this to True if you want to see the cart balancing the pole (slower)
render = False
#render = True
plt.ion()
env = CartPoleEnvironment()
if render:
renderer = CartPoleRenderer()
env.setRenderer(renderer)
renderer.start()
module = ActionValueNetwork(4, 3)
task = DiscreteBalanceTask(env, 100)
learner = NFQ()
learner.explorer.epsilon = 0.4
agent = LearningAgent(module, learner)
testagent = LearningAgent(module, None)
experiment = EpisodicExperiment(task, agent)
def plotPerformance(values, fig):
plt.figure(fig.number)
plt.clf()
plt.plot(values, 'o-')
plt.gcf().canvas.draw()
from pybrain.rl.agents import LearningAgent
from pybrain.rl.learners.valuebased import NFQ, ActionValueNetwork
from pybrain.rl.experiments import EpisodicExperiment
from environment import Environment
from tasks import BalanceTask
from training import NFQTraining
task = BalanceTask()
action_value_function = ActionValueNetwork(task.outdim, task.nactions,
name='BalanceNFQActionValueNetwork')
learner = NFQ()
learner.gamma = 0.9999
learner.explorer.epsilon = 0.9
task.discount = learner.gamma
agent = LearningAgent(action_value_function, learner)
performance_agent = LearningAgent(action_value_function, None)
experiment = EpisodicExperiment(task, agent)
tr = NFQTraining('balance_nfq', experiment, performance_agent)
tr.train(7000, performance_interval=1, n_performance_episodes=1, plotsave_interval=10, plot_action_history=True)
if len( sys.argv ) < 2 or ( int( sys.argv[1] ) < 0 or int( sys.argv[1] ) > 4 ):
print 'Must supply a model type:\n\t1 = Uncert&Salience, 2 = Salience, 3 = Uncert, 4 = Activation!'
sys.exit()
if len( sys.argv ) < 3:
print 'Must supply an output file!'
sys.exit()
type = int( sys.argv[1] ) # 1 = Uncert&Salience, 2 = Salience, 3 = Uncert, 4 = Activation
env = DistractorRatio() # Create an instance of the D-R task
# Create an action/value neural net with an state space of 100 and an action space of 8
if type == 1:
module = ActionValueNetwork( 99, 7 )
else:
module = ActionValueNetwork( 51, 4 )
learner = NFQ()
learner.offPolicy = False # Disable off policy learning
#learner.explorer = HumanExplorer()
learner.explorer.epsilon = 0.4
#learner.explorer.decay = 0.99
agent = HumanAgent( module, learner, type ) # Create an agent that learns with NFQ
testagent = HumanAgent( module, None, type ) # Create a testing agent
experiment = CustomEpisodicExperiment( env, agent ) # Put the agent in the environment
if len( sys.argv ) == 4:
print 'Loading saved net...'
module.network = NetworkReader.readFrom( sys.argv[3] )
from scipy import *
import sys, time
from pybrain.rl.learners.valuebased import ActionValueNetwork
from pybrain.rl.agents import LearningAgent
from pybrain.rl.learners import Q, SARSA, NFQ
from pybrain.rl.experiments import Experiment
from pybrain.rl.environments import Task
from tasktest import TestTask
from envtest import TestEnv
env = TestEnv()
task = TestTask(env)
controller = ActionValueNetwork(1, 2)
learner = NFQ()
agent = LearningAgent(controller, learner)
experiment = Experiment(task, agent)
while True:
experiment.doInteractions(100)
agent.learn()
agent.reset()
print("Cycle")
def q_learning_nfq(**args):
# estimate
best_score = 0
best_turn = 1000
best_agent = None
score_list = []
turn_list = []
#for i in range(2):
for i in range(50):
# agent 初期化
# rand = 1.0
# learner = NFQ(maxEpochs=100)
# learner._setExplorer(EpsilonGreedyExplorer(rand))
controller = ActionValueNetwork(12, 4)
learner = NFQ()
agent = LearningAgent(controller, learner)
# training
print
print "==========================="
print 'before training'
print_state(agent.module.getValue)
training(agent, args)
print 'after training'
print_state(agent.module.getValue)
agent.learner._setExplorer(EpsilonGreedyExplorer(0.3))
score, turn = play(agent, 'neural', args, [2, 2])
score_list.append(score)
turn_list.append(turn)
print
print i, int(numpy.mean(score_list)), max(score_list), score, turn
# if i % args['episodes'] == 0:
# try:
# agent.learn()
# except:
# pass
# finally:
# agent.reset()
# # rand = fit_greedy(i)
# # agent.learner._setExplorer(EpsilonGreedyExplorer(rand))
# # if not i == 0 :
# # import sys
# # sys.exit()
# print i, int(numpy.mean(score_list)) , max(score_list) , score, turn
if best_score < score or best_turn > turn:
best_score = score
best_turn = turn
best_agent = agent
with open(args['path'] + '/result.dump', 'w') as f:
pickle.dump([score_list, turn_list, best_agent], f)
print
print "==========================="
print 'best score : ', best_score
print 'best turn : ', best_turn
print_state(agent.module.getValue)
