def main(_):
params = {}
if FLAGS.symbolic:
params = {'seed': FLAGS.seed, 'level_name': FLAGS.level_name}
env_generator = symbolic_alchemy.get_symbolic_alchemy_level
else:
env_settings = dm_alchemy.EnvironmentSettings(
seed=FLAGS.seed, level_name=FLAGS.level_name)
params = {'name': FLAGS.docker_image_name, 'settings': env_settings}
env_generator = dm_alchemy.load_from_docker
with env_generator(**params) as env:
agent = RandomAgent(env.action_spec())
timestep = env.reset()
score = 0
while not timestep.last():
action = agent.act(timestep)
timestep = env.step(action)
if timestep.reward:
score += timestep.reward
print('Total score: {:.2f}, reward: {:.2f}'.format(
score, timestep.reward))
def run(env_name, agent_name, nb_episodes, render_freq, render_mode):
logger.set_level(logger.INFO)
env = gym.make(env_name)
# You provide the directory to write to (can be an existing
# directory, including one with existing data -- all monitor files
# will be namespaced). You can also dump to a tempdir if you'd
# like: tempfile.mkdtemp().
#outdir = '/tmp/random-agent-results'
#video_callable = None if render_mode == 'human' else False
#env = wrappers.Monitor(env, directory=outdir, force=True, video_callable=video_callable)
#env = DynamicMonitor(env, directory=outdir, force=True, video_callable=video_callable)
env.render(mode=render_mode)
env.seed(0)
if agent_name == 'RandomAgent':
agent = RandomAgent(env.env.action_space)
elif agent_name == 'EpsilonGreedyAgent':
agent = EpsilonGreedy(env.env.action_space)
elif agent_name == 'GradientBanditAgent':
agent = GradientBandit(env.env.action_space)
elif agent_name == 'ucb':
agent = ucb(env.env.action_space)
elif agent_name == 'ThompsonSampling':
agent = ThompsonSampling(env.env.action_space)
step = 0
reward = 0
done = False
for episode in range(nb_episodes):
print(f'--------- Episode {episode} ---------')
ob = env.reset()
agent = agent.reset()
while True:
step += 1
# action space may have change
# agent = EpsilonGreedy(env.env.action_space)
action = agent.act(ob, reward, done)
ob, reward, done, _ = env.step(action)
if done:
break
if step % render_freq == 0:
env.render()
# Note there's no env.render() here. But the environment still can open window and
# render if asked by env.monitor: it calls env.render('rgb_array') to record video.
# Video is not recorded every episode, see capped_cubic_video_schedule for details.
# Close the env and write monitor result info to disk
env.env.close()
def main(episode_count):
env = gym.make('CartPole-v0')
agent = RandomAgent(env.action_space.n)
for i in range(episode_count):
observation = env.reset() # initialize the environment
done = False
step = 0
while not done:
env.render()
action = agent.act(observation)
next_observation, reward, done, info = env.step(action)
if done:
print("Episode finished after {} timesteps".format(step + 1))
observation = next_observation
step += 1
class Learner(object):
def __init__(self, agent_type, tensorboard_dir, save_dir=None, player='offense', pretrained=None,
seed=1, episodes=20000, server_port=6000, max_steps=15000, save_freq=500, start=0):
if player == 'offense':
self.team = 'base_left'
self.mid_action = offense_mid_action
else:
self.team = 'base_right'
if player == 'goalie':
self.mid_action = goalie_mid_action
else:
self.mid_action = defense_mid_action
self.seed = seed
self.save_freq = save_freq
self.episodes = episodes
self.scale_actions = True
self.server_port = server_port
self.max_steps = max_steps
self.save_dir = os.path.join('./saved_models', save_dir)
self.tensorboard_dir = os.path.join('./tensorboard-log', tensorboard_dir)
self.pretrained = pretrained
self.start = start
self.env = make_env(player, self.server_port, self.team, self.scale_actions)
# configure redis
self.redis_instance = connect_redis()
# initialize teammate set
self.redis_instance.sadd('teammates', self.env.unum)
print('Number of teammates:', self.redis_instance.scard('teammates'))
if agent_type == 'PDDPG':
self.agent = PDDPGAgent(self.env.observation_space, self.env.action_space,
actor_kwargs={'hidden_layers': [1024, 512, 256, 256, 128, 128],
'init_type': "kaiming",
'init_std': 0.01, 'activation': 'leaky_relu'},
critic_kwargs={'hidden_layers': [1024, 512, 256, 256, 128, 128],
'init_type': "kaiming",
'init_std': 0.01, 'activation': 'leaky_relu'},
batch_size=32, # batch_size,
learning_rate_actor=0.001, # learning_rate_actor, # 0.0001
learning_rate_critic=0.001, # learning_rate_critic, # 0.001
gamma=0.99, # gamma, # 0.99
tau_actor=0.001, # tau,
tau_critic=0.001, # tau,
n_step_returns=True, # n_step_returns,
epsilon_steps=1000, # epsilon_steps,
epsilon_final=0.1, # epsilon_final,
replay_memory_size=500000, # replay_memory_size,
inverting_gradients=True, # inverting_gradients,
initial_memory_threshold=1000, # initial_memory_threshold,
beta=0.2, # beta,
clip_grad=1., # clip_grad,
use_ornstein_noise=False, # use_ornstein_noise,
adam_betas=(0.9, 0.999), # default 0.95,0.999
seed=self.seed)
elif agent_type == 'MAPDDPG':
self.agent = MAPDDPGAgent(self.env.observation_space, self.env.action_space,
actor_kwargs={'hidden_layers': [1024, 512, 256, 256, 128, 128],
'init_type': "kaiming",
'init_std': 0.01, 'activation': 'leaky_relu'},
critic_kwargs={'hidden_layers': [1024, 512, 256, 256, 128, 128],
'init_type': "kaiming",
'init_std': 0.01, 'activation': 'leaky_relu'},
batch_size=32, # batch_size,
learning_rate_actor=0.001, # learning_rate_actor, # 0.0001
learning_rate_critic=0.001, # learning_rate_critic, # 0.001
gamma=0.99, # gamma, # 0.99
tau_actor=0.001, # tau,
tau_critic=0.001, # tau,
n_step_returns=True, # n_step_returns,
epsilon_steps=1000, # epsilon_steps,
epsilon_final=0.1, # epsilon_final,
replay_memory_size=500000, # replay_memory_size,
inverting_gradients=True, # inverting_gradients,
initial_memory_threshold=1000, # initial_memory_threshold,
beta=0.2, # beta,
clip_grad=1., # clip_grad,
use_ornstein_noise=False, # use_ornstein_noise,
adam_betas=(0.9, 0.999), # default 0.95,0.999
seed=self.seed,
num_agents=int(self.redis_instance.get('num_agents')),
unum=self.env.unum)
elif agent_type == 'RANDOM':
self.agent = RandomAgent(observation_space=self.env.observation_space, action_space=self.env.action_space)
else:
raise NotImplementedError
self.writer = SummaryWriter(self.tensorboard_dir)
print('Agent:', self.agent)
def load_model(self, dir, i):
prefix = os.path.join(dir, str(i))
print('Evaluating model from', prefix, '...')
self.agent.actor.load_state_dict(torch.load(prefix + '_actor.pt', map_location='cpu'))
self.agent.critic.load_state_dict(torch.load(prefix + '_critic.pt', map_location='cpu'))
self.agent.actor.eval()
self.agent.critic.eval()
print('Models evaluated successfully')
def run(self):
# Random seed
# self.seed += 10000 * proc_id()
torch.manual_seed(self.seed)
np.random.seed(self.seed)
# later can sort by approximity (need to be same as in redis_manager)
self.all_agents = list(self.redis_instance.smembers('teammates'))
self.all_agents.sort()
# Prepare for interaction with environment
start_time = time.time()
# if isinstance(self.agent, RandomAgent):
# for i in range(self.episodes):
# obs = self.env.reset()
# obs = np.array(obs, dtype=np.float32, copy=False)
# print(obs)
#
# for j in range(self.local_steps_per_episode):
# act, act_param = self.agent.act(obs)
# action = mid_action(act, act_param)
#
# next_obs, reward, terminal, info = self.env.step(action)
# obs = next_obs
n_step_returns = True
update_ratio = 0.1
if self.save_freq > 0 and self.save_dir:
self.save_dir = os.path.join(self.save_dir, 'agent' + str(self.env.unum))
os.makedirs(self.save_dir, exist_ok=True)
if self.pretrained:
self.load_model(self.pretrained, self.start)
# train log
total_reward = 0.
returns = []
timesteps = []
goals = []
if isinstance(self.agent, PDDPGAgent) and not isinstance(self.agent, MAPDDPGAgent):
print('\n===========Start Training PDDPG===========')
# main loop
for i in range(self.start, self.start + self.episodes):
# save model
if self.save_freq > 0 and self.save_dir and (i + 1) % self.save_freq == 0:
prefix = os.path.join(self.save_dir, str(i + 1))
torch.save(self.agent.actor.state_dict(), prefix + '_actor.pt')
torch.save(self.agent.critic.state_dict(), prefix + '_critic.pt')
print('Models saved successfully at episode' + str(i + 1))
info = {'status': "NOT_SET"}
# initialize environment, reward and transitions
obs = self.env.reset()
obs = np.array(obs, dtype=np.float32, copy=False)
episode_reward = 0.
transitions = []
# get discrete action and continuous parameters
act, act_param, all_actions, all_action_parameters = self.agent.act(obs)
action = self.mid_action(act, act_param)
for j in range(self.max_steps):
next_obs, reward, terminal, info = self.env.step(action)
next_obs = np.array(next_obs, dtype=np.float32, copy=False)
# get discrete action and continuous parameters
next_act, next_act_param, next_all_actions, next_all_action_parameters = self.agent.act(next_obs)
next_action = self.mid_action(next_act, next_act_param)
if n_step_returns:
transitions.append(
[obs, np.concatenate((all_actions.data, all_action_parameters.data)).ravel(), reward,
next_obs, np.concatenate((next_all_actions.data,
next_all_action_parameters.data)).ravel(), terminal])
else:
self.agent.step(obs, (act, act_param, all_actions, all_action_parameters), reward, next_obs,
(next_act, next_act_param, next_all_actions, next_all_action_parameters),
terminal,
optimise=False)
act, act_param, all_actions, all_action_parameters = \
next_act, next_act_param, next_all_actions, next_all_action_parameters
action = next_action
obs = next_obs
episode_reward += reward
# env.render()
if terminal:
break
# decay epsilon
self.agent.end_episode()
# calculate n-step returns
if n_step_returns:
nsreturns = compute_n_step_returns(transitions, self.agent.gamma)
for t, nsr in zip(transitions, nsreturns):
t.append(nsr)
self.agent.replay_memory.append(state=t[0], action=t[1], reward=t[2], next_state=t[3],
next_action=t[4],
terminal=t[5], time_steps=None, n_step_return=nsr)
# update networks at the end of each episode
n_updates = int(update_ratio * j)
for _ in range(n_updates):
self.agent.update()
# train log
returns.append(episode_reward)
timesteps.append(j)
goals.append(info['status'] == 'GOAL')
total_reward += episode_reward
self.writer.add_scalar('Episode reward', episode_reward, i)
if i % 100 == 0:
print('{0:5s} : Total mean reward:{1:.4f} | Episode reward:{2:.4f}'
.format(str(i + 1), total_reward / (i + 1), episode_reward))
self.writer.add_scalar('Last 100episodes mean reward', np.array(returns[-100:]).mean(), i)
elif isinstance(self.agent, MAPDDPGAgent):
print('\n===========Start Training MAPDDPG===========')
# main loop
for i in range(self.start, self.start + self.episodes):
# save model
if self.save_freq > 0 and self.save_dir and (i + 1) % self.save_freq == 0:
prefix = os.path.join(self.save_dir, str(i + 1))
torch.save(self.agent.actor.state_dict(), prefix + '_actor.pt')
torch.save(self.agent.critic.state_dict(), prefix + '_critic.pt')
print('Models saved successfully at episode' + str(i + 1))
info = {'status': "NOT_SET"}
# initialize environment, reward and transitions
obs = self.env.reset()
obs = np.array(obs, dtype=np.float32, copy=False)
episode_reward = 0.
transitions = []
# get discrete action and continuous parameters
act, act_param, all_actions, all_action_parameters = self.agent.act(obs)
action = self.mid_action(act, act_param)
# update the observation and action of agent i in redis
sync_agent_obs_actions(self.redis_instance, self.env.unum, obs, all_actions, all_action_parameters)
# query all agents' observations and actions
all_agent_obs, all_agent_actions, success1 = query_all_obs_actions(self.redis_instance)
for j in range(self.max_steps):
# # TODO: query other agents' actions (for inference use)
# take action in environment
next_obs, reward, terminal, info = self.env.step(action)
next_obs = np.array(next_obs, dtype=np.float32, copy=False)
# get discrete action and continuous parameters
next_act, next_act_param, next_all_actions, next_all_action_parameters = self.agent.act(next_obs)
next_action = self.mid_action(next_act, next_act_param)
# update the observation and action of agent i in redis
sync_agent_obs_actions(self.redis_instance, self.env.unum, next_obs,
next_all_actions, next_all_action_parameters)
# query all agents' observations and actions
all_agent_next_obs, all_agent_next_actions, success2 = query_all_obs_actions(self.redis_instance)
if n_step_returns and success1 and success2:
transitions.append([all_agent_obs.ravel(), all_agent_actions.ravel(), reward,
all_agent_next_obs.ravel(), all_agent_next_actions.ravel(), terminal])
# else:
# self.agent.step(obs, (act, act_param, all_actions, all_action_parameters), reward, next_obs,
# (next_act, next_act_param, next_all_actions, next_all_action_parameters),
# terminal,
# optimise=False)
all_agent_actions = all_agent_next_actions
action = next_action
all_agent_obs = all_agent_next_obs
success1 = True
episode_reward += reward
if terminal:
break
# decay epsilon
self.agent.end_episode()
# calculate n-step returns
if n_step_returns:
nsreturns = compute_n_step_returns(transitions, self.agent.gamma)
for t, nsr in zip(transitions, nsreturns):
t.append(nsr)
if not any(elem is None for elem in t):
self.agent.replay_memory.append(state=t[0], action=t[1], reward=t[2], next_state=t[3],
next_action=t[4], terminal=t[5],
time_steps=None, n_step_return=nsr)
# update networks at the end of each episode
n_updates = int(update_ratio * j)
for _ in range(n_updates):
# sync policy in redis
sync_agent_policy(self.redis_instance, self.env.unum, self.agent.actor_target)
self.agent.update()
# train log
returns.append(episode_reward)
timesteps.append(j)
goals.append(info['status'] == 'GOAL')
total_reward += episode_reward
self.writer.add_scalar('Episode reward', episode_reward, i)
if i % 100 == 0:
print('{0:5s} : Total mean reward:{1:.4f} | Episode reward:{2:.4f}'
.format(str(i + 1), total_reward / (i + 1), episode_reward))
self.writer.add_scalar('Last 100episodes mean reward', np.array(returns[-100:]).mean(), i)
train_goal_ratio = goals.count(True) / self.episodes
print("Training goal ratio: %.2f" % train_goal_ratio)
end_time = time.time()
print("Training time: %.2f seconds" % (end_time - start_time))
print("==========TRAINING END==========")
agent = DistanceMinimizing(env.action_space)
elif agent_arg == 'deg-min':
agent = DegreeMinimizing(env.action_space)
seed = asteroids_args.seed if isinstance(asteroids_args.seed, int) else asteroids_args.seed[0]
runs = asteroids_args.runs if isinstance(asteroids_args.runs, int) else asteroids_args.runs[0]
env.seed(seed)
total_runs = 0
total_score = 0
while total_runs < runs:
reward = 0
done = False
score = 0
special_data = {'ale.lives': 3}
ob = env.reset()
i = 0
while not done:
action = agent.act(ob, reward, done)
ob, reward, done, x = env.step(action)
score += reward
# env.render()
i += 1
print(f'Asteroids score [{i}]: {score}')
total_runs += 1
total_score += score
print(f'Average score across {total_runs} runs: {total_score / total_runs}')
# Close the env and write monitor result info to disk
env.close()
import gym
from agents.random_agent import RandomAgent
env = gym.make("CartPole-v0")
agent = RandomAgent(env.action_space)
for i_episode in range(20):
observation = env.reset()
reward = 0
done = False
for t in range(100):
env.render()
print(observation)
action = agent.act(object, reward, done)
observation, reward, done, info = env.step(action)
if done:
print("Episode finished after {} timesteps".format(t + 1))
break
env.close()