RLax (pronounced "relax") is a library built on top of JAX that exposes useful building blocks for implementing reinforcement learning agents. Full documentation can be found at rlax.readthedocs.io.
You can install the latest released version of RLax from PyPI via:
pip install rlaxor you can install the latest development version from GitHub:
pip install git+https://github.com/deepmind/rlax.gitAll RLax code may then be just in time compiled for different hardware
(e.g. CPU, GPU, TPU) using jax.jit.
In order to run the examples/ you will also need to clone the repo and
install the additional requirements:
optax,
haiku, and
bsuite.
The operations and functions provided are not complete algorithms, but implementations of reinforcement learning specific mathematical operations that are needed when building fully-functional agents capable of learning:
- Values, including both state and action-values;
- Values for Non-linear generalizations of the Bellman equations.
- Return Distributions, aka distributional value functions;
- General Value Functions, for cumulants other than the main reward;
- Policies, via policy-gradients in both continuous and discrete action spaces.
The library supports both on-policy and off-policy learning (i.e. learning from data sampled from a policy different from the agent's policy).
See file-level and function-level doc-strings for the documentation of these functions and for references to the papers that introduced and/or used them.
See examples/ for examples of using some of the functions in RLax to
implement a few simple reinforcement learning agents, and demonstrate learning
on BSuite's version of the Catch environment (a common unit-test for
agent development in the reinforcement learning literature):
Other examples of JAX reinforcement learning agents using rlax can be found in
bsuite.
Reinforcement learning studies the problem of a learning system (the agent), which must learn to interact with the universe it is embedded in (the environment).
Agent and environment interact on discrete steps. On each step the agent selects an action, and is provided in return a (partial) snapshot of the state of the environment (the observation), and a scalar feedback signal (the reward).
The behaviour of the agent is characterized by a probability distribution over actions, conditioned on past observations of the environment (the policy). The agents seeks a policy that, from any given step, maximises the discounted cumulative reward that will be collected from that point onwards (the return).
Often the agent policy or the environment dynamics itself are stochastic. In this case the return is a random variable, and the optimal agent's policy is typically more precisely specified as a policy that maximises the expectation of the return (the value), under the agent's and environment's stochasticity.
There are three prototypical families of reinforcement learning algorithms:
- those that estimate the value of states and actions, and infer a policy by inspection (e.g. by selecting the action with highest estimated value)
- those that learn a model of the environment (capable of predicting the observations and rewards) and infer a policy via planning.
- those that parameterize a policy that can be directly executed,
In any case, policies, values or models are just functions. In deep reinforcement learning such functions are represented by a neural network. In this setting, it is common to formulate reinforcement learning updates as differentiable pseudo-loss functions (analogously to (un-)supervised learning). Under automatic differentiation, the original update rule is recovered.
Note however, that in particular, the updates are only valid if the input data is sampled in the correct manner. For example, a policy gradient loss is only valid if the input trajectory is an unbiased sample from the current policy; i.e. the data are on-policy. The library cannot check or enforce such constraints. Links to papers describing how each operation is used are however provided in the functions' doc-strings.
We define functions and operations for agents interacting with a single stream
of experience. The JAX construct vmap can be used to apply these same
functions to batches (e.g. to support replay and parallel data generation).
Many functions consider policies, actions, rewards, values, in consecutive
timesteps in order to compute their outputs. In this case the suffix _t and
tm1 is often to clarify on which step each input was generated, e.g:
q_tm1: the action value in thesourcestate of a transition.a_tm1: the action that was selected in thesourcestate.r_t: the resulting rewards collected in thedestinationstate.discount_t: thediscountassociated with a transition.q_t: the action values in thedestinationstate.
Extensive testing is provided for each function. All tests should also verify
the output of rlax functions when compiled to XLA using jax.jit and when
performing batch operations using jax.vmap.
This repository is part of the DeepMind JAX Ecosystem, to cite Rlax please use the citation:
@software{deepmind2020jax,
title = {The {D}eep{M}ind {JAX} {E}cosystem},
author = {DeepMind and Babuschkin, Igor and Baumli, Kate and Bell, Alison and Bhupatiraju, Surya and Bruce, Jake and Buchlovsky, Peter and Budden, David and Cai, Trevor and Clark, Aidan and Danihelka, Ivo and Dedieu, Antoine and Fantacci, Claudio and Godwin, Jonathan and Jones, Chris and Hemsley, Ross and Hennigan, Tom and Hessel, Matteo and Hou, Shaobo and Kapturowski, Steven and Keck, Thomas and Kemaev, Iurii and King, Michael and Kunesch, Markus and Martens, Lena and Merzic, Hamza and Mikulik, Vladimir and Norman, Tamara and Papamakarios, George and Quan, John and Ring, Roman and Ruiz, Francisco and Sanchez, Alvaro and Sartran, Laurent and Schneider, Rosalia and Sezener, Eren and Spencer, Stephen and Srinivasan, Srivatsan and Stanojevi\'{c}, Milo\v{s} and Stokowiec, Wojciech and Wang, Luyu and Zhou, Guangyao and Viola, Fabio},
url = {http://github.com/deepmind},
year = {2020},
}
