# feats=FeatureStack([RandFourierFeat(env.obs_space.flat_dim, num_feat=20, bandwidth=env.obs_space.bound_up)])
feats=FeatureStack([const_feat, identity_feat, sign_feat, abs_feat, squared_feat, cubic_feat, ATan2Feat(1, 2),
MultFeat([3, 4])])
)
policy = LinearPolicy(spec=env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
max_iter=500,
eps=1.0,
pop_size=20*policy.num_param,
num_rollouts=4,
expl_std_init=0.2,
expl_std_min=0.02,
use_map=True,
optim_mode='scipy',
num_workers=12,
)
algo = REPS(ex_dir, env, policy, **algo_hparam)
# Save the hyper-parameters
save_list_of_dicts_to_yaml([
dict(env=env_hparams, seed=args.seed),
dict(policy=policy_hparam),
dict(algo=algo_hparam, algo_name=algo.name)],
ex_dir
)
# Jeeeha
algo.train(snapshot_mode='best', seed=args.seed)
# feats=FeatureStack(RBFFeat(num_feat_per_dim=20, bounds=env.obs_space.bounds, scale=0.8)),
feats=FeatureStack(identity_feat, sin_feat))
policy = LinearPolicy(spec=env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
max_iter=500,
eps=0.2,
pop_size=10 * policy.num_param,
num_init_states_per_domain=10,
expl_std_init=0.2,
expl_std_min=0.02,
num_epoch_dual=1000,
optim_mode="scipy",
lr_dual=1e-3,
use_map=True,
num_workers=8,
)
algo = REPS(ex_dir, env, policy, **algo_hparam)
# Save the hyper-parameters
save_dicts_to_yaml(
dict(env=env_hparams, seed=args.seed),
dict(policy=policy_hparam),
dict(algo=algo_hparam, algo_name=algo.name),
save_dir=ex_dir,
)
# Jeeeha
algo.train(seed=args.seed)
cubic_feat,
ATan2Feat(1, 2),
MultFeat((3, 4)),
]))
policy = LinearPolicy(spec=env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
max_iter=500,
eps=1.0,
pop_size=20 * policy.num_param,
num_init_states_per_domain=4,
expl_std_init=0.2,
expl_std_min=0.02,
use_map=True,
optim_mode="scipy",
num_workers=12,
)
algo = REPS(ex_dir, env, policy, **algo_hparam)
# Save the hyper-parameters
save_dicts_to_yaml(
dict(env=env_hparams, seed=args.seed),
dict(policy=policy_hparam),
dict(algo=algo_hparam, algo_name=algo.name),
save_dir=ex_dir,
)
# Jeeeha
algo.train(snapshot_mode="best", seed=args.seed)