def __init__(self, batch_size, max_size, beta):
self.batch_size = batch_size # mini batch大小
self.max_size = 2**math.floor(
math.log2(max_size)) # 保证 sum tree 为完全二叉树
self.beta = beta
self._sum_tree = SumTree(max_size)
def __init__(self, buffer_size, seed):
"""Initialize a ReplayBuffer object.
Params
======
seed (int): random seed
"""
self.memory = SumTree(buffer_size)
self.experience = namedtuple("Experience", field_names=["state", "action", "reward", "next_state", "done"])
self.seed = random.seed(seed)
# epsilon: small amount to avoid zero priority
# alpha: [0~1] determines how much prioritization is used. with 0, we would get the uniform case
# beta: Controls importance-sampling compensation. fully compensates for the non-uniform probabilities
# when beta=1. The unbiased nature of the updates is most important near convergence at the end of
# training, so we define a schedule on the exponent beta that starts from initial value and reaches 1
# only at the end of learning.
self.epsilon = 0.01
self.alpha = 0.6
beta_start = 0.4
self.beta_end = 1.0
self.beta = beta_start
beta_increments = 200
self.beta_increment = (self.beta_end - beta_start)/beta_increments
def __init__(self,
buffer_size,
alpha,
beta_zero,
beta_increment_size=0.001,
epsilon=0.1,
max_priority=1.,
seed=None):
"""Priority replay buffer initialiser.
Args:
buffer_size (int): capacity of the replay buffer.
alpha (float): priority scaling hyperparameter.
beta_zero (float): importance sampling scaling hyperparameter.
beta_increment_size (float): beta annealing rate.
epsilon (float): base priority to ensure non-zero sampling probability.
max_priority (float): initial maximum priority.
seed (int): seed for random number generator
"""
random.seed(seed)
self.sum_tree = SumTree(buffer_size)
self.memory = {}
self.experience = namedtuple(
"experience", ["state", "action", "reward", "next_state", "done"])
self.buffer_size = buffer_size
self.beta_increment_size = beta_increment_size
self.max_priority = max_priority**alpha
self.min_priority = max_priority**alpha
self.last_min_update = 0
self.alpha = alpha
self.beta = beta_zero
self.epsilon = epsilon
def test_add(self):
instance = SumTree(4)
instance.add(p=1, data=1)
np.testing.assert_array_equal([1, 1, 0, 1, 0, 0, 0], instance.tree)
instance.add(p=2, data=2)
np.testing.assert_array_equal([3, 3, 0, 1, 2, 0, 0], instance.tree)
def __init__(self, max_size, alpha, eps):
self.max_size = max_size
self.alpha = alpha
self.eps = eps
self.tree = SumTree(max_size)
self.last_idxs = None
self.size = 0
def __init__(self, capacity, batch_size):
self.capacity = capacity
self.batch_size = batch_size
self.tree = SumTree(capacity=capacity)
self.alpha = 0.6
self.beta = 0.4
self.p_epsilon = 1e-4
self.batch_size = 50
def __init__(self, memory_size):
self.memory_size = memory_size
self.memory = SumTree(memory_size)
self.epsilon = 0.0001 # small amount to avoid zero priority
self.alpha = 0.6 # adj_pri = pri^alpha
self.beta = 0.4 # importance-sampling, from initial value increasing to 1
self.beta_max = 1
self.beta_increment_per_sampling = 0.001
self.abs_err_upper = 1. # clipped td error
def __init__(self, action_size, buffer_size, batch_size, alpha, epsilon):
self.action_size = action_size
self.tree = SumTree(buffer_size)
self.batch_size = batch_size
self.experience = namedtuple(
"Experience",
field_names=["state", "action", "reward", "next_state", "done"])
self.alpha = alpha
self.epsilon = epsilon
def __init__(self, capacity):
"""
Instantiate a priority based memory with capable of holding
capacity experiences. Memories are sampled with frequency
based on their priority.
"""
# Circular buffer array based tree with priorities as node values.
self.tree = SumTree(capacity)
self.e = 0.01 # Small constant to ensure all priorities > 0
self.a = 0.6 # Constant to control the weight of error on priority
def __init__(self, e, a, beta, beta_increment_per_sampling, capacity,
max_priority):
self.capacity = capacity
self.e = e
self.a = a
self.beta = beta
self.beta_increment_per_sampling = beta_increment_per_sampling
self.capacity = capacity
self.max_priority = max_priority
self.sum_tree = SumTree(self.capacity)
def __init__(self):
self.limit = MEMORY_CAPACITY
self.err_tree = SumTree(MEMORY_CAPACITY)
self.action_shape = (0, MEMORY_ACTION_CNT)
self.reward_shape = (0, MEMORY_REWARD_CNT)
self.terminal_shape = self.action_shape
self.observation_shape = (0, MEMORY_CRITIC_FEATURE_NUM)
self.store_times = 0
self.Transition = namedtuple(
'Transition',
('state', 'action', 'reward', 'next_state', 'terminal'))
def __init__(self, alpha, beta, beta_end, epsilon, num_steps, replay_size):
self.alpha = alpha
self.beta_start = beta
self.beta_end = beta_end
self.beta = beta
self.epsilon = epsilon
self.num_steps = num_steps
self.memory = SumTree(replay_size)
self.replay_size = replay_size
def __init__(self,
observation_len: int,
action_len: int,
reward_len: int,
capacity: int,
alpha: int = 0.6):
super(PriorityBuffer, self).__init__(observation_len, action_len,
reward_len, capacity)
self.sum_tree = SumTree(capacity)
self.max_priority = alpha
self.min_priority = alpha
self.alpha = alpha
def __init__(self,
tree_memory_length,
error_multiplier=0.01,
alpha=0.6,
beta=0.4,
beta_increment_per_sample=0.001):
self.tree = SumTree(tree_memory_length)
self.tree_memory_length = tree_memory_length
self.error_multiplier = error_multiplier
self.per_alpha = alpha
self.per_beta_init = beta
self.beta_increment_per_sample = beta_increment_per_sample
def __init__(self,
capacity,
alpha=0.6,
beta=0.4,
beta_anneal_step=0.001,
epsilon=0.00000001):
self.tree = SumTree(capacity)
self.capacity = capacity
self.a = alpha
self.beta = beta
self.beta_increment_per_sampling = beta_anneal_step
self.e = epsilon
def load(self, lst_serializable):
"""
Load pickable representation of Replay Buffer. Inverse function of serializable
"""
super().load(lst_serializable[0])
self.max_priority = lst_serializable[1][0]
self.min_priority = lst_serializable[1][1]
self.alpha = lst_serializable[1][2]
capacity = lst_serializable[1][3]
tree_index = range(capacity)
self.sum_tree = SumTree(capacity)
self.sum_tree.update_values(tree_index, lst_serializable[1][4])
def __init__(self, host_name, db_name, collection_name):
self.host_name = host_name
self.db_name = db_name
self.collection_name = collection_name
self.client = MongoClient(host_name, 27017)
self.db = self.client[db_name]
self.replay_memory_collection = self.db[collection_name]
self.sum_tree = SumTree(self.capacity)
memory_priorities = self.replay_memory_collection.find({},
{"priority": 1})
for memory_priority in memory_priorities:
self.sum_tree.add(memory_priority["priority"],
{"_id": memory_priority["_id"]})
def test_len(self):
instance = SumTree(4)
instance.add(p=1, data=1)
self.assertEqual(len(instance), 1)
instance.add(p=2, data=2)
self.assertEqual(len(instance), 2)
instance.add(p=3, data=3)
instance.add(p=4, data=4)
instance.add(p=5, data=5)
self.assertEqual(len(instance), 4)
def __init__(self, params):
buffer_size = params['buffer_size']
batch_size = params['batch_size']
mode = params['mode']
self.__buffer_size = buffer_size
self.__batch_size = batch_size
self.__mode = mode
self.__experience = namedtuple(
"Experience",
field_names=["state", "action", "reward", "next_state", "done"])
self.__memory = SumTree(buffer_size)
self.__memory_buffer = []
def __init__(self, action_size, buffer_size, batch_size, seed):
"""Initialize a ReplayBuffer object.
Params
======
action_size (int): dimension of each action
buffer_size (int): maximum size of buffer
batch_size (int): size of each training batch
seed (int): random seed
"""
self.action_size = action_size
self.memory = SumTree(buffer_size)
self.batch_size = batch_size
self.experience = namedtuple("Experience", field_names=["state", "action", "reward", "next_state", "done"])
self.seed = random.seed(seed)
def __init__(self, device, memory_size, update_every=4, seed=0):
""" Initializes the data structure
:param device: (torch.device) Object representing the device where to allocate tensors
:param memory_size: (int) Maximum capacity of memory buffer
:param update_every: (int) Number of steps between update operations
:param seed: (int) Seed used for PRNG
"""
self.device = device
self.probability_weights = SumTree(capacity=memory_size, seed=seed)
self.elements = deque(maxlen=memory_size)
self.update_every = update_every
self.step = 0
self.experience = namedtuple("Experience", field_names=["state", "action", "reward", "next_state", "done"])
def __init__(self,
capacity,
state_size=37,
epsilon=0.001,
alpha=0.4,
beta=0.3,
beta_increment_per_sampling=0.001,
abs_err_upper=1):
self.tree = SumTree(capacity)
self.epsilon = epsilon # Avoid 0 priority and hence a do not give a chance for the priority to be selected stochastically
self.alpha = alpha # Vary priority vs randomness. alpha = 0 pure uniform randomnes. Alpha = 1, pure priority
self.beta = beta # importance-weight-sampling, from small to big to give more importance to corrections done towards the end of the training
self.beta_increment_per_sampling = 0.001
self.abs_err_upper = 1 # clipped abs error
self.state_size = state_size
def __init__(
self,
buffer_size,
batch_size,
seed,
beta_start=0.4,
delta_beta=1e-5,
alpha=0.6,
eps=1e-8,
):
"""Initialize PER.
Args:
buffer_size (int): Size of replay buffer. The actual size will be the
first power of 2 greater than buffer_size.
batch_size (int): Size of batches to draw.
seed (float): Seed.
beta_start (float): Initial value for beta (importance sampling exponent)
delta_beta (float): Beta increment at each time step.
alpha (float): Priority exponent.
eps (float): Small positive number to avoid unsampling 0 prioritized examples.
"""
# Depth of sum tree
depth = int(math.log2(buffer_size)) + 1
super(PrioritizeReplayBuffer, self).__init__(2**depth, batch_size,
seed)
# Initialize sum tree to keep track of the sum of priorities
self.priorities = SumTree(depth)
# Current max priority
self.max_p = 1.0
# PER Parameters
self.alpha = alpha
self.eps = eps
self.beta = beta_start
self.delta_beta = delta_beta
def __init__(self, action_size, agent_config):
"""Initialize a PriorityBuffer object.
Params
======
action_size (int): dimension of each action
buffer_size (int): maximum size of buffer
batch_size (int): size of each training batch
seed (int): random seed
a (float): amount of uniformity in the sampling (0 == uniform, 1. == priority only)
beta_start (float): start of beta value for prioritised buffer
beta_max_steps (int): max number of steps to reach beta value of 1.
"""
self.action_size = action_size
self.tree = SumTree(capacity=agent_config.buffer_size)
self.batch_size = agent_config.batch_size
# self.seed = random.seed(buffer_config.seed)
self.epsilon = agent_config.buffer_epsilon
# how much randomness we require a = 0 (pure random) a = 1 (only priority)
self.alpha = agent_config.alpha
self.beta = agent_config.beta_start
self.beta_start = agent_config.beta_start
self.beta_end = agent_config.beta_end
self.beta_increment_per_sampling = (self.beta_end - self.beta_start) / agent_config.beta_max_steps
def __init__(self, capacity):
self.tree = SumTree(capacity)
self.capacity = capacity
def __init__(self):
global MEMORY_LEN
self.tree = SumTree(MEMORY_LEN)
def create_tree(sample):
tree = SumTree(len(sample))
for e in sample:
tree.add(p=e, data=e)
return tree
def __init__(self, buffer_size, alpha):
self.capacity = buffer_size
self.tree = SumTree(buffer_size)
self.alpha = alpha
self.max_priority = 1
def __init__(self, config, capacity):
self.config = config
self.capacity = capacity
self.tree = SumTree(capacity)
def __init__(self, obs_dim, act_dim, size, act_dtype):
SimpleMemory.__init__(self, obs_dim, act_dim, size, act_dtype)
self.tree = SumTree(size)
self.tree_lock = Lock()
