def experiment(variant):
env = variant['env_class']()
env = normalize(env)
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
epoch_discount_schedule_class = variant['epoch_discount_schedule_class']
epoch_discount_schedule = epoch_discount_schedule_class(
**variant['epoch_discount_schedule_params'])
algorithm = DDPG(env,
exploration_strategy=es,
qf=qf,
policy=policy,
epoch_discount_schedule=epoch_discount_schedule,
**variant['algo_params'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
# env = HalfCheetahEnv()
# env = PointEnv()
env = gym_env("Pendulum-v0")
# env = HopperEnv()
horizon = variant['algo_params']['max_path_length']
env = TimeLimitedEnv(env, horizon)
env = normalize(env)
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
algorithm = MultiStepDdpg(env,
exploration_strategy=es,
qf=qf,
policy=policy,
**variant['algo_params'])
algorithm.train()
return algorithm.final_score
def example(variant):
env = CartpoleEnv()
env = NormalizedBoxEnv(env)
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
**variant['qf_params'],
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = DDPG(
env,
qf,
policy,
exploration_policy,
**variant['algo_params']
)
algorithm.to(ptu.device)
algorithm.train()
def example(variant):
env = HalfCheetahEnv()
if variant['normalize']:
env = normalize(env)
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
32,
32,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
32,
32,
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = DDPG(env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
**variant['algo_params'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = variant['env_class'](**variant['env_params'])
env = normalize(env)
es = OUStrategy(
action_space=env.action_space,
**variant['es_params']
)
algo_class = variant['algo_class']
algo_params = variant['algo_params']
hidden_size = variant['hidden_size']
if algo_class == DDPG:
# algo_params.pop('naf_policy_learning_rate')
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
hidden_size,
hidden_size,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
hidden_size,
hidden_size,
)
algorithm = DDPG(
env,
exploration_strategy=es,
qf=qf,
policy=policy,
**variant['algo_params']
)
elif algo_class == NAF:
algo_params.pop('qf_learning_rate')
# algo_params.pop('policy_learning_rate')
qf = NafPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
hidden_size,
)
algorithm = NAF(
env,
policy=qf,
exploration_strategy=es,
**variant['algo_params']
)
else:
raise Exception("Invalid algo class: {}".format(algo_class))
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = NormalizedBoxEnv(MultiGoalEnv(
actuation_cost_coeff=10,
distance_cost_coeff=1,
goal_reward=10,
))
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
100,
100,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
100,
100,
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
plotter = QFPolicyPlotter(
qf=qf,
# policy=policy,
policy=exploration_policy,
obs_lst=np.array([[-2.5, 0.0],
[0.0, 0.0],
[2.5, 2.5]]),
default_action=[np.nan, np.nan],
n_samples=100
)
algorithm = DDPG(
env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
render_eval_paths=True,
plotter=plotter,
**variant['algo_params']
)
algorithm.to(ptu.device)
algorithm.train()
def example(variant):
env = variant['env_class']()
env = normalize(env)
es = OUStrategy(action_space=env.action_space)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
policy = FeedForwardPolicy(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
400,
300,
)
algorithm = DDPG(env,
exploration_strategy=es,
qf=qf,
policy=policy,
**variant['algo_params'])
algorithm.to(ptu.device)
algorithm.train()