class Environment:
def __init__(self, num_items, use_all=False, val=False):
self.env = PointingEnv(num_items=num_items, use_all=use_all, val=val)
def run(self, agent, inspect=False):
s = self.env.reset()
R = 0
global failed_cnt
global done_cnt
global R_total
while True:
sbar = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
a = agent.act(sbar)
#print a
s_, r, done = self.env.step(a)
s = s_
R += r
if done:
if r < 1:
failed_cnt += 1
else:
done_cnt += 1
R_total += R
break
class Environment:
def __init__(self, num_items, use_all=False, val=False):
self.env = PointingEnv(num_items=num_items, use_all=use_all, val=val)
def run(self, agent):
s = self.env.reset()
R = 0
global r_history
imaginary = False #flag to start using imaginary rollouts for training
if episodes > I_START and USE_IMAGINARY == True:
#print('Imaginary training started...')
imaginary = True #start using imaginary rollouts
# TODO: decide if imaginary or not
while True:
z, sigma = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
a = agent.act(z)
# print('a = ', a)
if imaginary:
for i_b in range(I_B):
agent.brain.env_model.init_model(z)
zhat = z
for i_d in range(I_D):
zhat_, rhat, donehat = agent.brain.env_model.step(a)
rhat, donehat = round(rhat), round(donehat)
#if donehat == 1:
# zhat_ = None
agent.observe((zhat, a, rhat, zhat_, donehat),
imaginary=True)
if donehat == 1:
break
zhat = zhat_
s_, r, done = self.env.step(a)
z_, sigma_ = agent.brain.encode(
np.reshape(s_, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
#if done: # terminal state
# z_ = None
agent.observe((z, a, r, z_, done, sigma, sigma_), imaginary=False)
if (episodes >= ENV_LEARN_START) and USE_IMAGINARY and ENV_LEARN:
for i in range(R_ENV):
agent.train_env()
for i in range(R_CR):
agent.replay(imaginary=False)
if imaginary:
for i in range(R_CI):
agent.replay(imaginary=imaginary)
s = s_
R += r
r_history[episodes] = R
if done:
break
print("Total reward:", R, ", episode: ", episodes)
class Environment:
def __init__(self, num_items):
self.env = PointingEnv(num_items)
self.env_model = EnvironmentModel()
def run(self, agent):
s = self.env.reset()
sbar = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
R = 0
imaginary = False
if episodes > 0:
imaginary = True
self.env_model.init_model(sbar)
#TODO: decide if imaginary or not
while True:
a = agent.act(sbar)
#print('a = ', a)
if imaginary:
sbar_, r, done = self.env_model.step(a)
r, done = round(r), round(done)
else:
s_, r, done = self.env.step(a)
sbar_ = agent.brain.encode(
np.reshape(s_, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
if done: # terminal state
sbar_ = None
agent.observe((sbar, a, r, sbar_, done), imaginary=imaginary)
agent.replay(imaginary=imaginary)
if imaginary:
sbar = sbar_
else:
s = s_
sbar = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
R += r
if done:
break
print("Total reward:", R, ", episode: ", episodes)
class Environment:
def __init__(self, num_items):
self.env = PointingEnv(num_items)
def run(self, agent, inspect=False):
s = self.env.reset()
R = 0
global selected
global ss
global r_history
if not selected:
ss = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
selected = True
# print (agent.brain.predictOne(ss))
while True:
if inspect: self.env.printState()
sbar = agent.brain.encode(
np.reshape(s, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
# print(sbar)
a = agent.act(sbar)
# print(a)
s_, r, done = self.env.step(a)
sbar_ = agent.brain.encode(
np.reshape(s_, (1, IMAGE_WIDTH, IMAGE_HEIGHT, CHANNELS)))
if done: # terminal state
sbar_ = None
agent.observe((sbar, a, r, sbar_, done))
agent.replay()
s = s_
R += r
r_history[episodes] = R
if done:
break
print("Total reward:", R, ", episode: ", episodes)
def __init__(self, num_items, use_all=False, val=False):
self.env = PointingEnv(num_items=num_items, use_all=use_all, val=val)
def __init__(self, num_items):
self.env = PointingEnv(num_items)
self.env_model = EnvironmentModel()
def __init__(self, num_items):
self.env = PointingEnv(num_items)