def build_credit_env(config=None, initial_state=None):
"""Construct credit environment that is parameterized by the initial state.
This allows the user to specify different datasets other than the default
Credit dataset.
"""
if config is None:
config = Config()
elif not isinstance(config, Config):
raise ValueError(f"Config must be an instance of {type(Config())}")
if initial_state is None:
initial_state = State()
elif not isinstance(initial_state, State):
raise ValueError(
f"Initial state must be an instance of {type(State())}")
config.base_state = copy.deepcopy(initial_state)
return ODEEnvBuilder(
simulate_fn=simulate,
config=config,
# The initial state is the baseline features and labels in the credit dataset
initial_state=initial_state,
# Action space is classifiers with the same number of parameters are
# features.
action_space=credit_action_space(initial_state),
# Observation space is the strategically adapted features and labels
observation_space=credit_observation_space(initial_state),
timestep=1,
intervention_fn=compute_intervention,
reward_fn=compute_reward,
)
def observation_space():
"""Return the model observation space."""
num_states = State.num_variables()
state_space_low = np.zeros(num_states)
state_space_high = np.inf * np.ones(num_states)
return spaces.Box(state_space_low, state_space_high, dtype=np.float64)
World3Env = ODEEnvBuilder(
simulate_fn=simulate,
# Smaller delta_t improves numerical stability
config=Config(delta_t=0.5),
initial_state=State(),
# In this environment there are 9 actions defined by
# nonrenewable_resource_usage and pollution_generation_factor.
action_space=spaces.Discrete(9),
observation_space=observation_space(),
timestep=1.0,
intervention_fn=get_intervention,
reward_fn=get_reward,
)
register(
id="world3-v0",
entry_point=World3Env,
max_episode_steps=400,
reward_threshold=1e5,
)
CreditEnv = ODEEnvBuilder(
simulate_fn=simulate,
config=Config(),
# The initial state is the baseline features and labels in the credit dataset
initial_state=State(features=CreditData.features,
labels=CreditData.labels),
# Action space is classifiers with the same number of parameters are
# features.
action_space=spaces.Box(low=-np.inf,
high=np.inf,
shape=(CreditData.num_features, ),
dtype=np.float64),
# Observation space is the strategically adapted features and labels
observation_space=spaces.Dict({
"features":
spaces.Box(
low=-np.inf,
high=np.inf,
shape=CreditData.features.shape,
dtype=np.float64,
),
"labels":
spaces.Box(
low=-np.inf,
high=np.inf,
shape=CreditData.labels.shape,
dtype=np.float64,
),
}),
timestep=1,
intervention_fn=compute_intervention,
reward_fn=compute_reward,
)
# lost economic output per time
reward -= economic_output_per_time * current_social_distancing
return reward
def observation_space():
"""Return observation space.
The state is (susceptible, exposed, infected, recovered).
"""
state_dim = State.num_variables()
state_space_low = np.zeros(state_dim)
state_space_high = np.inf * np.ones(state_dim)
return spaces.Box(state_space_low, state_space_high, dtype=np.float64)
Covid19Env = ODEEnvBuilder(
simulate_fn=simulate,
config=Config(),
initial_state=State(),
action_space=spaces.Discrete(6),
observation_space=observation_space(),
timestep=1.0,
intervention_fn=get_intervention,
reward_fn=get_reward,
)
register(
id="COVID19-POOR-v0", entry_point=Covid19Env, max_episode_steps=150,
reward_threshold=1e10,
)
def observation_space():
"""Return the observation space. The state is (nonmedical_users, oud_useres, illicit_users)."""
state_dim = State.num_variables()
state_space_low = np.zeros(state_dim)
state_space_high = np.inf * np.ones(state_dim)
return spaces.Box(state_space_low, state_space_high, dtype=np.float64)
OpioidEnv = ODEEnvBuilder(
simulate_fn=simulate,
config=Config(),
initial_state=State(),
# In this environment we define 4 actions:
# - Do nothing
# - Reduce nonmedical opioid use by 5%
# - Reduce illicit opioid use by 5%
# - Reduce both by 5%
action_space=spaces.Discrete(4),
observation_space=observation_space(),
timestep=1.0,
intervention_fn=get_intervention,
reward_fn=get_reward,
)
register(
id="opioid-v0",
entry_point=OpioidEnv,
# The simulator starts in 2002 and ends in 2030.
max_episode_steps=28,
reward_threshold=0,
)
The state is (uninfected_T1, infected_T1, uninfected_T2, infected_T2,
free_virus, immune_response) in units (cells/ml, cells/ml, cells/ml,
cells/ml, copies/ml, cells/ml).
"""
state_dim = State.num_variables()
state_space_low = np.zeros(state_dim)
state_space_high = np.inf * np.ones(state_dim)
return spaces.Box(state_space_low, state_space_high, dtype=np.float64)
HivEnv = ODEEnvBuilder(
simulate_fn=simulate,
config=Config(),
initial_state=State(),
# In this environment there are 4 actions defined by
# epsilon_1 = 0 or 0.7 and epsilon_2 = 0 or 0.3.
action_space=spaces.Discrete(4),
observation_space=observation_space(),
timestep=1.0,
intervention_fn=get_intervention,
reward_fn=get_reward,
)
register(
id="HIV-v0",
entry_point=HivEnv,
max_episode_steps=400,
reward_threshold=1e10,
)