def __init__(self,
actions,
gamma=0.9,
r_max=1.,
v_max=None,
deduce_v_max=True,
n_known=None,
deduce_n_known=True,
epsilon_q=0.1,
epsilon_m=None,
delta=None,
n_states=None,
min_sampling_probability=0.1,
name="MaxQInit"):
"""
:param actions: action space of the environment
:param gamma: (float) discount factor
:param r_max: (float) known upper-bound on the reward function
:param v_max: (float) known upper-bound on the value function
:param deduce_v_max: (bool) set to True to deduce v_max from r_max
:param n_known: (int) count after which a state-action pair is considered known
(only set n_known if delta and epsilon are not defined)
:param deduce_n_known: (bool) set to True to deduce n_known from (delta, n_states, epsilon_m)
:param epsilon_q: (float) precision of value iteration algorithm for Q-value computation
:param epsilon_m: (float) precision of the learned models in L1 norm
:param delta: (float) models are learned epsilon_m-closely with probability at least 1 - delta
:param n_states: (int) number of states
:param min_sampling_probability: (float) minimum sampling probability of an environment
:param name: (str)
"""
RMax.__init__(self,
actions=actions,
gamma=gamma,
r_max=r_max,
v_max=v_max,
deduce_v_max=deduce_v_max,
n_known=n_known,
deduce_n_known=deduce_n_known,
epsilon_q=epsilon_q,
epsilon_m=epsilon_m,
delta=delta,
n_states=n_states,
name=name)
self.min_sampling_probability = min_sampling_probability
self.SA_memory = defaultdict(lambda: defaultdict(lambda: False))
self.U_memory = [] # Upper-bounds on the Q-values of previous MDPs
self.n_required_tasks = number_of_tasks_for_high_confidence_upper_bound(
delta, min_sampling_probability)
def __init__(self,
actions,
gamma=.9,
r_max=1.,
v_max=None,
deduce_v_max=True,
n_known=None,
deduce_n_known=True,
epsilon_q=0.1,
epsilon_m=None,
delta=None,
n_states=None,
name="ExpRMax",
path='results/'):
RMax.__init__(self,
actions=actions,
gamma=gamma,
r_max=r_max,
v_max=v_max,
deduce_v_max=deduce_v_max,
n_known=n_known,
deduce_n_known=deduce_n_known,
epsilon_q=epsilon_q,
epsilon_m=epsilon_m,
delta=delta,
n_states=n_states,
name=name)
# Recorded variables
self.discounted_return = 0.
self.total_return = 0.
self.n_time_steps = 0 # nb of time steps
self.update_time_steps = [] # time steps where a model update occurred
self.path = path
self.instance_number = 0
self.run_number = 0
def __init__(self,
actions,
gamma=.9,
r_max=1.,
v_max=None,
deduce_v_max=True,
n_known=None,
deduce_n_known=True,
epsilon_q=0.1,
epsilon_m=None,
delta=None,
n_states=None,
max_memory_size=None,
prior=None,
estimate_distances_online=True,
min_sampling_probability=.1,
name="LRMax"):
"""
:param actions: action space of the environment
:param gamma: (float) discount factor
:param r_max: (float) known upper-bound on the reward function
:param v_max: (float) known upper-bound on the value function
:param deduce_v_max: (bool) set to True to deduce v_max from r_max
:param n_known: (int) count after which a state-action pair is considered known
:param deduce_n_known: (bool) set to True to deduce n_known from (delta, n_states, epsilon_m)
(only set n_known if delta and epsilon are not defined)
:param epsilon_q: (float) precision of value iteration algorithm for Q-value computation
:param epsilon_m: (float) precision of the learned models in L1 norm
:param delta: (float) models are learned epsilon_m-closely with probability at least 1 - delta
:param n_states: (int) number of states
:param max_memory_size: (int) maximum number of saved models (infinity if None)
:param prior: (float) prior knowledge of maximum model's distance
:param estimate_distances_online: (bool) set to True for online estimation of a tighter upper-bound for the
model pseudo-distances. The estimation is valid with high probability.
:param min_sampling_probability: (float) minimum sampling probability of an environment
:param name: (str)
"""
self.name = name
RMax.__init__(self,
actions=actions,
gamma=gamma,
r_max=r_max,
v_max=v_max,
deduce_v_max=deduce_v_max,
n_known=n_known,
deduce_n_known=deduce_n_known,
epsilon_q=epsilon_q,
epsilon_m=epsilon_m,
delta=delta,
n_states=n_states,
name=name)
# Lifelong Learning memories
self.max_memory_size = max_memory_size
self.U_memory = []
self.R_memory = []
self.T_memory = []
self.SA_memory = defaultdict(lambda: defaultdict(lambda: False))
self.U_lip = []
self.b = self.epsilon_m * (1. + self.gamma * self.v_max)
# Prior knowledge on maximum model distance
prior_max = self.r_max + self.gamma * 2. * self.v_max
self.prior = prior_max if prior is None else min(prior, prior_max)
self.prior = round(self.prior, 2)
# Online distances estimation
self.estimate_distances_online = estimate_distances_online
self.min_sampling_probability = min_sampling_probability
self.D = defaultdict(lambda: defaultdict(lambda: prior_max)
) # Dictionary of distances (high probability)
self.n_samples_high_confidence = compute_n_samples_high_confidence(
min_sampling_probability, delta)
self.update_upper_bound()