def experiment_explore_vs_exploit(meta_path, save_directory, setups, episodes, steps):
alpha = 0.0001
basis_order = 3
env = gym.make('CartPole-v0')
env.reset()
(obs, reward, done, info) = env.step(env.action_space.sample())
obs = EnvWrapper.modified_sigmoid(obs)
phi = fourier_basis(obs, order=basis_order)
num_features = phi.shape[0]
num_actions = env.action_space.n
if os.path.isdir(save_directory) == False:
os.mkdir(save_directory)
num_samples = 5
meta = pickle.load(open(meta_path, "rb"))
agents = []
for setup in setups:
gym_env = gym.make('CartPole-v0')
gym_env.env.force_mag = setup["force"]
gym_env.env.length = setup["pole_length"]
gym_env.env.masscart = setup["masscart"]
gym_env.env.masspole = setup["masspole"]
env = EnvWrapper(gym_env, basis_order=basis_order, normalization=0)
agent = LinearAgent(env, meta_policy=meta, alpha=alpha, algo="SARSA")
agents.append( agent )
policies = []
for agent in agents:
rewards = agent.train(num_episodes=episodes, max_steps=steps, verbose=True, update_meta=False, render=False)
policies.append( copy.deepcopy(agent.learning_algorithm) )
rewards = []
for i, agent in enumerate(agents):
agent.learning_algorithm = policies[i]
agent.random_action_prob = 0.0
agent.RANDOM_ACTION_DECAY = 1.0
exploit_rewards = agent.train(num_episodes=episodes, max_steps=steps, verbose=True, update_meta=False, render=False)
agent.random_action_prob = 1.0
explore_rewards = agent.train(num_episodes=episodes, max_steps=steps, verbose=True, update_meta=False, render=False)
rewards.append( {"explore" : explore_rewards, "exploit" : exploit_rewards} )
pickle.dump(rewards, open(save_directory+"/explore_exploit.pkl", "wb"))
import time
from PIL import Image
if __name__ == "__main__":
usage = "usage: %prog [options] arg"
parser = OptionParser(usage)
parser.add_option("-e", "--environment", action="store", help="environment name", type="string", default="animat")
parser.add_option("-m", "--model", action="store", help="env model", type="string")
parser.add_option("-a", "--actor", action="store", help="agent actor", type="string")
parser.add_option("-c", "--critic", action="store", help="agent critic", type="string")
parser.add_option("-s", "--save_path", action="store", help="path for saving vid", type="string", default=None)
(options, args) = parser.parse_args()
env = options.environment
if env != "animat":
env = gym.make( env )
env.env.model_xml = options.model
else:
gym_env = AnimatEnv(options.model)
env = EnvWrapper(gym_env, basis_order=3, normalization=1)
agent = LinearAgent(env, meta_policy=None, algo="PPO")
agent.random_action_prob = 0.0
agent.learning_algorithm.load_model(options.actor, options.critic)
agent.play( max_steps=10000, delay=0.01, save_path=options.save_path)
time.sleep(0.5)
def experiment_explore_vs_exploit(meta_path, save_directory, setups,
episodes, steps):
alpha = 0.001
basis_order = 3
env = AnimatEnv("./CustomEnvironments/maze1.txt")
env.reset()
(obs, reward, done, info) = env.step(env.action_space.sample())
obs = EnvWrapper.normalize_range(obs, env.env_range)
phi = fourier_basis(obs, order=basis_order)
num_features = phi.shape[0]
num_actions = env.action_space.n
if os.path.isdir(save_directory) == False:
os.mkdir(save_directory)
num_samples = 5
meta = pickle.load(open(meta_path, "rb"))
agents = []
for setup in setups:
gym_env = AnimatEnv(setup)
env = EnvWrapper(gym_env, basis_order=basis_order, normalization=1)
agent = LinearAgent(env,
meta_policy=meta,
alpha=alpha,
algo="REINFORCE")
agents.append(agent)
policies = []
for agent in agents:
rewards = agent.train(num_episodes=episodes,
max_steps=steps,
verbose=True,
update_meta=False,
render=False)
policies.append(copy.deepcopy(agent.learning_algorithm))
rewards = []
for i, agent in enumerate(agents):
agent.learning_algorithm = policies[i]
agent.random_action_prob = 0.0
agent.RANDOM_ACTION_DECAY = 1.0
exploit_rewards = agent.train(num_episodes=episodes,
max_steps=steps,
verbose=True,
update_meta=False,
render=False)
agent.random_action_prob = 1.0
explore_rewards = agent.train(num_episodes=episodes,
max_steps=steps,
verbose=True,
update_meta=False,
render=False)
rewards.append({
"explore": explore_rewards,
"exploit": exploit_rewards
})
pickle.dump(rewards, open(save_directory + "/explore_exploit.pkl",
"wb"))
def experiment_meta_vs_random(meta_actor, meta_critic, save_directory, setups, episodes, steps):
alpha = 0.001
basis_order = 3
env = gym.make('CartPole-v0')
env.reset()
(obs, reward, done, info) = env.step(env.action_space.sample())
obs = EnvWrapper.modified_sigmoid(obs)
phi = fourier_basis(obs, order=basis_order)
num_features = phi.shape[0]
num_actions = env.action_space.n
if os.path.isdir(save_directory) == False:
os.mkdir(save_directory)
num_samples = 5
for k in range(2):
if k == 0:
meta = MetaPolicy(num_features=num_features, num_actions=num_actions, algo="PPO", alpha=1e-3, beta=1e-3, env=env)
meta.learning_algorithm.load_model(meta_actor, meta_critic)
else:
meta = None
agents = []
for setup in setups:
domain_agents = []
for _ in range(num_samples):
gym_env = gym.make('CartPole-v0')
gym_env.env.force_mag = setup["force"]
gym_env.env.length = setup["pole_length"]
gym_env.env.masscart = setup["masscart"]
gym_env.env.masspole = setup["masspole"]
env = EnvWrapper(gym_env, basis_order=basis_order, normalization=0)
agent = LinearAgent(env, meta_policy=meta, alpha=alpha, algo="PPO")
if meta is None:
agent.random_action_prob = 0.0
domain_agents.append( agent )
agents.append( domain_agents )
domain_rewards_by_episode = {}
for ep in range(episodes):
trajectories_by_domain = ExperimentsCartpole._run_episode(domain_agents=agents, num_steps=steps, optimize_meta=False)
domain_samples = []
for i in trajectories_by_domain.keys():
sample_rewards = []
for j in range(len(trajectories_by_domain[i])):
t_rewards = []
for t in trajectories_by_domain[i][j]:
t_rewards.append( t['reward'] )
sample_rewards.append( t_rewards )
domain_samples.append( sample_rewards )
print("Episode %d" %(ep))
domain_rewards_by_episode[ep] = domain_samples
if ep % 100 == 0:
filename = "meta_test"+str(ep)+".pkl" if k == 0 else "no_meta_test_"+str(ep)+".pkl"
pickle.dump(domain_rewards_by_episode, open(save_directory+"/"+filename, "wb"))
def experiment_meta_vs_random(meta_actor, meta_critic, save_directory,
xml_models, episodes, steps):
alpha = 1e-4
beta = 1e-3
env = gym.make("RoboschoolHopper-v1")
env.reset()
(obs, reward, done, info) = env.step(env.action_space.sample())
num_features = obs.shape[0]
num_actions = env.action_space.low.shape[0]
if os.path.isdir(save_directory) == False:
os.mkdir(save_directory)
num_samples = 5
for k in range(2):
if k == 0:
meta = MetaPolicy(num_features=num_features,
num_actions=num_actions,
algo="PPO",
alpha=1e-3,
beta=1e-3,
env=env)
meta.learning_algorithm.load_model(meta_actor, meta_critic)
else:
meta = None
agents = []
for model in xml_models:
domain_agents = []
for _ in range(num_samples):
env = gym.make("RoboschoolHopper-v1")
env.env.model_xml = model
agent = LinearAgent(env,
meta_policy=meta,
algo="PPO",
alpha=alpha,
beta=beta)
agent.learning_algorithm.t_length = 128
agent.learning_algorithm.update_steps = 128
agent.learning_algorithm.epochs = 8
if meta is None:
agent.random_action_prob = 0.0
domain_agents.append(agent)
agents.append(domain_agents)
domain_rewards_by_episode = {}
for ep in range(episodes):
null_actions = {}
trajectories_by_domain = ExperimentsHopper._run_episode(
domain_agents=agents, num_steps=steps, optimize_meta=False)
domain_samples = []
for i in trajectories_by_domain.keys():
sample_returns = []
for j in range(len(trajectories_by_domain[i])):
sample_returns.append(sum(
trajectories_by_domain[i][j]))
domain_samples.append(sample_returns)
print("Episode %d / %d" % (ep, episodes))
domain_rewards_by_episode[ep] = domain_samples
if ep % 100 == 0:
filename = "meta_test_" + str(
ep) + ".pkl" if k == 0 else "no_meta_test_" + str(
ep) + ".pkl"
pickle.dump(domain_rewards_by_episode,
open(save_directory + "/" + filename, "wb"))
for ai, a in enumerate(agents):
type_ = "meta_" if k == 0 else "no_meta_"
type_ += str(ai) + "_"
a[0].learning_algorithm.save_model(
save_directory + "/" + type_, ep)