replaytables-andnp 6.2.16

A simple replay buffer implementation in python for sampling n-step trajectories

ReplayTables

Benchmarks

Getting started

Installation:

pip install ReplayTables-andnp

Basic usage:

from typing import NamedTuple
from ReplayTables.ReplayBuffer import ReplayBuffer

class Data(NamedTuple):
    x: np.ndarray
    a: np.ndarray
    r: np.ndarray

buffer = ReplayBuffer(
    max_size=100_000,
    structure=Data,
    rng=np.random.default_rng(0),
)

buffer.add(Data(x, a, r))

batch = buffer.sample(32)
print(batch.x.shape) # -> (32, d)
print(batch.a.shape) # -> (32, )
print(batch.r.shape) # -> (32, )

Prioritized Replay

An implementation of prioritized experience replay from

Schaul, Tom, et al. "Prioritized experience replay." ICLR (2016).

The defaults for this implementation strictly adhere to the defaults from the original work, though several configuration options are available.

from typing import NamedTuple
from ReplayTables.PER import PERConfig, PrioritizedReplay

class Data(NamedTuple):
    a: float
    b: float

# all configurables are optional.
config = PERConfig(
    # can also use "mean" mode to place new samples in the middle of the distribution
    # or "given" mode, which requires giving the priority when the sample is added
    new_priority_mode='max',
    # the sampling distribution is a mixture between uniform sampling and the priority
    # distribution. This specifies the weight given to the uniform sampler.
    # Setting to 1 reverts this back to an inefficient form of standard uniform replay.
    uniform_probability=1e-3,
    # this implementation assume priorities are positive. Can scale priorities by raising to
    # some power. Default is `priority**(1/2)`
    priority_exponent=0.5,
    # if `new_priority_mode` is 'max', then the buffer tracks the highest seen priority.
    # this can cause accidental saturation if outlier priorities are observed. This provides
    # an exponential decay of the max in order to prevent permanent saturation.
    max_decay=1,
)

# if no config is given, defaults to original PER parameters
buffer = PrioritizedReplay(
    max_size=100_000,
    structure=Data,
    rng=np.random.default_rng(0),
    config=config,
)

buffer.add(Data(a=1, b=2))

# if `new_priority_mode` is 'given':
buffer.add(Data(a=1, b=2), priority=1.3)

batch = buffer.sample(32)