def main():
options, args = make_optparse().parse_args()
print "Loading data"
data = load_dataset(options.datapath)
print "Building network"
net = rnn.RecurrentNetwork(74, options.n_hidden, 1, outfunc='sig')
print "Number of parameters:", len(net.parameters)
objfunc = make_objective_func(net, data, rnn.SumOfSquares())
x0 = scipy.random.standard_normal(len(net.parameters)) * 0.1
optimizer = PGPE(objfunc, x0, minimize=True)
optimizer.maxEvaluations = options.maxevals
print "First fitness:", objfunc(x0)
print "Optimizing..."
params, fitness = optimizer.learn()
print "Last fitness:", fitness
true_positives, total = stats(net, data)
print "Total positives found: %i (%.2f)" % (true_positives, true_positives /
total)
return 0
def main():
options, args = make_optparse().parse_args()
print "Loading data"
data = load_dataset(options.datapath)
print "Building network"
net = rnn.RecurrentNetwork(74, options.n_hidden, 1, outfunc='sig')
print "Number of parameters:", len(net.parameters)
objfunc = make_objective_func(net, data, rnn.SumOfSquares())
x0 = scipy.random.standard_normal(len(net.parameters)) * 0.1
optimizer = PGPE(objfunc, x0, minimize=True)
optimizer.maxEvaluations = options.maxevals
print "First fitness:", objfunc(x0)
print "Optimizing..."
params, fitness = optimizer.learn()
print "Last fitness:", fitness
true_positives, total = stats(net, data)
print "Total positives found: %i (%.2f)" % (true_positives,
true_positives / total)
return 0
env = None
for runs in range(numbExp):
# create environment
#Options: Bool(OpenGL), Bool(Realtime simu. while client is connected), ServerIP(default:localhost), Port(default:21560)
if env != None: env.closeSocket()
env = ShipSteeringEnvironment()
# create task
task = GoNorthwardTask(env, maxsteps=500)
# create controller network
net = buildNetwork(task.outdim, task.indim, outclass=TanhLayer)
# create agent with controller and learner (and its options)
agent = OptimizationAgent(
net,
PGPE(learningRate=0.3,
sigmaLearningRate=0.15,
momentum=0.0,
epsilon=2.0,
rprop=False,
storeAllEvaluations=True))
et.agent = agent
#create experiment
experiment = EpisodicExperiment(task, agent)
#Do the experiment
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch)
et.printResults((agent.learner._allEvaluations)[-50:-1], runs, updates)
et.addExps()
et.showExps()
#To view what the simulation is doing at the moment set the environment with True, go to pybrain/rl/environments/ode/ and start viewer.py (python-openGL musst be installed, see PyBrain documentation)
exp_desciptions = []
for i in range(num_experiments):
# set up environment, task, neural net, agent, and experiment
env = InfoMaxEnv(object_names, action_names, num_objects, False)
task = InfoMaxTask(env, max_steps=max_steps)
net = buildNetwork(task.outdim,
task.indim,
bias=True,
outclass=SoftmaxLayer)
if algorithm == 'pgpe':
agent = OptimizationAgent(
net,
PGPE(storeAllEvaluations=True, minimize=False, verbose=False))
elif algorithm == 'cmaes':
agent = OptimizationAgent(net, CMAES(minimize=False,
verbose=False))
experiment = EpisodicExperiment(task, agent)
exp_desciptions.append([i, agent, experiment, task])
pool = Pool(processes=num_cpus)
res = []
if algorithm == 'pgpe':
res = pool.map(run_experiment, exp_desciptions)
elif algorithm == 'cmaes':
for desc in exp_desciptions:
def main():
if len(sys.argv) != 2:
print 'Please provide a path to a model data directory.'
print ('The script will load the newest model data from the directory,'
'then continue to improve that model')
sys.exit(0)
model_directory = sys.argv[1]
existing_models = sorted(glob(os.path.join(model_directory, '*.rlmdl')))
if existing_models:
newest_model_name = existing_models[-1]
iteration_count = int(newest_model_name[-12:-6]) + 1
print 'Loading model {}'.format(newest_model_name)
newest_model = open(newest_model_name, 'r')
agent = pickle.load(newest_model)
else:
net = buildNetwork(Environment.outdim,
Environment.outdim + Environment.indim,
Environment.indim)
agent = OptimizationAgent(net, PGPE())
iteration_count = 1
environment = Environment(LOCAL_HOST, PORT, PATH_TO_SCENE)
task = Task(environment)
experiment = EpisodicExperiment(task, agent)
def signal_handler(signal, frame):
print 'Exiting gracefully'
environment.teardown()
sys.exit(0)
signal.signal(signal.SIGINT, signal_handler)
while True:
time.sleep(1)
print '>>>>> Running iteration {}'.format(iteration_count)
# NOTE this weird stuff is hacky, but we need it to plug in our autosave
# stuff properly. Took a long time to figure this out.
experiment.optimizer.maxEvaluations = experiment.optimizer.numEvaluations + experiment.optimizer.batchSize
try:
experiment.doEpisodes()
except Exception as e:
print 'ERROR RUNNING SIMULATION: \n{}'.format(e)
environment.teardown()
else:
if iteration_count % AUTOSAVE_INTERVAL == 0:
filename = str(iteration_count).zfill(6) + '.rlmdl'
filename = os.path.join(model_directory, filename)
f = open(filename, 'w+')
print 'Saving model to {}'.format(filename)
pickle.dump(agent, f)
iteration_count += 1
print 'Iteration finished <<<<<'
def PGPE_Factory():
return PGPE()
numbExp = 10 #number of experiments
et = ExTools(batch, prnts) #tool for printing and plotting
env = None
for runs in range(numbExp):
# create environment
#Options: Bool(OpenGL), Bool(Realtime simu. while client is connected), ServerIP(default:localhost), Port(default:21560)
if env != None: env.closeSocket()
env = JohnnieEnvironment()
# create task
task = StandingTask(env)
# create controller network
net = buildNetwork(len(task.getObservation()),
hiddenUnits,
env.actLen,
outclass=TanhLayer)
# create agent with controller and learner (and its options)
agent = OptimizationAgent(net, PGPE(storeAllEvaluations=True))
et.agent = agent
# create the experiment
experiment = EpisodicExperiment(task, agent)
#Do the experiment
for updates in range(epis):
for i in range(prnts):
experiment.doEpisodes(batch)
et.printResults((agent.learner._allEvaluations)[-50:-1], runs, updates)
et.addExps()
et.showExps()
#To view what the simulation is doing at the moment, go to pybrain/rl/environments/ode/ and start viewer.py (python-openGL musst be installed, see PyBrain documentation)