def run(self):
super(MarioEnv, self).run()
self.env = gym_super_mario_bros.make(self.env_id)
self.env = BinarySpaceToDiscreteSpaceEnv(self.env, SIMPLE_MOVEMENT)
self.reset()
print('[ Worker %2d ] ' % (self.idx), end='')
print('Playing <', self.env_id, '>')
self.request_action(0, False)
while True:
action = self.child_conn.recv()
next_state, reward, done, info = self.env.step(action)
self.steps += 1
self.accum_reward += reward
next_state = rgb2dataset(next_state)
if self.is_render and self.idx == 0:
self.env.render()
# make a transition
self.transition.append(next_state)
if len(self.transition) > 4:
self.transition.pop(0)
if done:
self.send_result(info['x_pos'])
self.reset()
self.request_action(reward, True)
else:
self.request_action(reward, False)
def run(self):
super(MarioEnv, self).run()
self.model = A3C(
self.s_dim,
self.a_dim,
gamma=0.95,
epsilon_start=1.0,
epsilon_end=0.1,
epsilon_length=100000,
use_cuda=self.use_cuda,
)
self.model.l_net.load_state_dict(self.g_net.state_dict())
self.env = gym_super_mario_bros.make(self.env_id)
self.env = BinarySpaceToDiscreteSpaceEnv(self.env, SIMPLE_MOVEMENT)
self.reset()
print('[ Worker %2d ] ' % (self.idx), end='')
print('Playing <', self.env_id, '>')
while True:
if len(self.transition) != 4:
action = self.model.get_action(self.transition, is_random=True)
else:
action = self.model.get_action(self.transition,
is_random=False)
next_state, reward, done, info = self.env.step(action)
self.steps += 1
self.accum_reward += reward
next_state = rgb2dataset(next_state)
if self.is_render and self.idx == 0:
self.env.render()
self.buffer_state.append(self.transition)
self.buffer_action.append(action)
self.buffer_reward.append(reward)
if len(self.buffer_state
) > 0 and self.steps % self.update_iter == 0:
next_transition = self.transition[1:]
next_transition.append(next_state)
self.train(next_transition, done)
self.buffer_state.clear()
self.buffer_action.clear()
self.buffer_reward.clear()
# make a transition
self.transition.append(next_state)
if len(self.transition) > 4:
self.transition.pop(0)
if done:
self.send_result(info['x_pos'])
self.reset()
def reset(self):
state = self.env.reset()
state = rgb2dataset(state)
self.transition.clear()
self.transition.append(state)
self.steps = 0
self.episodes += 1
self.accum_reward = 0
def reset(self):
state = self.env.reset()
state = rgb2dataset(state)
self.transition = np.zeros([4, 84, 84])
self.transition[-1, :] = state
self.steps = 0
self.episodes += 1
self.accum_reward = 0
def run(self):
super(MarioEnv, self).run()
self.env = gym_super_mario_bros.make(self.env_id)
self.env = BinarySpaceToDiscreteSpaceEnv(self.env, SIMPLE_MOVEMENT)
self.reset()
print('[ Worker %2d ] ' % (self.idx), end='')
print('Playing <', self.env_id, '>')
self.request_action(0, False)
while True:
action = self.child_conn.recv()
# print(SIMPLE_MOVEMENT[action])
next_state, reward, done, info = self.env.step(action)
force_done = False
if reward == -15:
force_done = True
self.steps += 1
self.accum_reward += reward
next_state = rgb2dataset(next_state)
if self.is_render and self.idx == 0:
self.env.render()
# make a transition
self.transition[:3, :, :] = self.transition[1:, :, :]
self.transition[3, :, :] = next_state
if done:
self.send_result(self.prev_xpos)
self.reset()
self.request_action(reward, force_done)
else:
self.request_action(reward, force_done)
self.prev_xpos = info['x_pos']
epsilon_end=0.01,
epsilon_length=300000,
use_cuda=use_cuda,
lr=lr,
replay_buffer_size=replay_buffer_size,
train_start_step=train_start_step,
batch_size=batch_size,
target_update_interval=target_update_interval,
train_step_interval=train_step_interval)
writer = SummaryWriter(log_dir='runs/ddqn')
while model.episode < max_episode:
state = env.reset()
state = rgb2dataset(state)
model.episode += 1
accum_reward = 0
# Transition
transition = []
transition.append(state)
while True:
if len(transition) == 4:
action = model.get_action(transition, is_random=False)
else:
action = model.get_action(transition, is_random=True)
state_, reward, done, info = env.step(action)