def someEpisodes(game_env, net, discountFactor=0.99, maxSteps=100, avgOver=1, returnEvents=False, exploretoo=True):
""" Return the fitness value for one episode of play, given the policy defined by a neural network. """
import pdb
pdb.set_trace()
task = GameTask(game_env)
game_env.recordingEnabled = True
game_env.reset()
net.reset()
task.maxSteps=maxSteps
agent = LearningAgent(net)
agent.learning = False
agent.logging = False
exper = EpisodicExperiment(task, agent)
fitness = 0
for _ in range(avgOver):
rs = exper.doEpisodes(1)
# add a slight bonus for more exploration, if rewards are identical
if exploretoo:
fitness += len(set(game_env._allEvents)) * 1e-6
# the true, discounted reward
fitness += sum([sum([v*discountFactor**step for step, v in enumerate(r)]) for r in rs])
fitness /= avgOver
if returnEvents:
return fitness, game_env._allEvents
else:
return fitness
def someEpisodes(game_env, net, discountFactor=0.99, maxSteps=100, avgOver=1, returnEvents=False):
""" Return the fitness value for one episode of play, given the policy defined by a neural network. """
task = GameTask(game_env)
game_env.recordingEnabled = True
game_env.reset()
net.reset()
task.maxSteps=maxSteps
agent = LearningAgent(net)
agent.learning = False
agent.logging = False
exper = EpisodicExperiment(task, agent)
fitness = 0
for _ in range(avgOver):
rs = exper.doEpisodes(1)
# add a slight bonus for more exploration, if rewards are identical
fitness += len(set(game_env._allEvents)) * 1e-6
# the true, discounted reward
fitness += sum([sum([v*discountFactor**step for step, v in enumerate(r)]) for r in rs])
fitness /= avgOver
if returnEvents:
return fitness, game_env._allEvents
else:
return fitness
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
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
experiment = Experiment(task, agent)
k = 0
while k < number_of_runs:
experiment.doInteractions(96)
agent.learn()
agent.reset()
#log some data of the first and last run.
if k == 0: # if it is the first run
first_run_time2 = environment.log_time
first_run_soc2 = environment.log_soc
first_run_volt2 = environment.log_volt
#if k == number_of_runs - 1: # if it it the last run
# last_run_time2 = environment.log_time
# last_run_soc2 = environment.log_soc
# last_run_volt2 = environment.log_volt
environment.reset()
k += 1
agent.learning = False #to keep it from exploring
experiment = Experiment(task, agent)
experiment.doInteractions(96)
last_run_time2 = environment.log_time
last_run_soc2 = environment.log_soc
last_run_volt2 = environment.log_volt
plt.plot(first_run_time2, first_run_soc2, 'r--', last_run_time2, last_run_soc2,
'g--')
plt.show()
