from ifqi.models.actionregressor import ActionRegressor
from ifqi.models.regressor import Regressor
from ifqi.models.mlp import MLP
from ifqi.models.ensemble import Ensemble
"""
Simple script to quickly run fqi. It solves the Acrobot environment according
to the experiment presented in:
Ernst, Damien, Pierre Geurts, and Louis Wehenkel.
"Tree-based batch mode reinforcement learning."
Journal of Machine Learning Research 6.Apr (2005): 503-556.
"""
mdp = envs.CarOnHill()
state_dim, action_dim, reward_dim = envs.get_space_info(mdp)
assert reward_dim == 1
regressor_params = {'n_estimators': 50,
'criterion': 'mse',
'min_samples_split': 5,
'min_samples_leaf': 2,
'input_scaled': False,
'output_scaled': False}
discrete_actions = mdp.action_space.values
# ExtraTrees
regressor = Regressor(ExtraTreesRegressor, **regressor_params)
# Action regressor of Ensemble of ExtraTreesEnsemble
# regressor = Ensemble(ExtraTreesRegressor, **regressor_params)
regressor = ActionRegressor(regressor, discrete_actions=discrete_actions,
from ifqi import envs
from ifqi.evaluation import evaluation
from ifqi.evaluation.utils import check_dataset, split_data_for_fqi
from ifqi.models.regressor import Regressor
from ifqi.models.mlp import MLP
from ifqi.models.linear import Ridge
from ifqi.algorithms.pbo.pbo import PBO
"""
Simple script to quickly run pbo. It solves the LQG environment.
"""
mdp = envs.LQG1D()
state_dim, action_dim, reward_dim = envs.get_space_info(mdp)
reward_idx = state_dim + action_dim
discrete_actions = np.linspace(-8, 8, 20)
dataset = evaluation.collect_episodes(mdp, n_episodes=100)
check_dataset(dataset, state_dim, action_dim, reward_dim)
sast, r = split_data_for_fqi(dataset, state_dim, action_dim, reward_dim)
### Q REGRESSOR ##########################
class LQG_Q():
def __init__(self):
self.w = np.array([1., 0.])
def predict(self, sa):
k, b = self.w
#print(k,b)
return - b * b * sa[:, 0] * sa[:, 1] - 0.5 * k * sa[:, 1] ** 2 - 0.4 * k * sa[:, 0] ** 2