Dopamax

Dopamax is a library containing pure JAX implementations of common reinforcement learning algorithms. Everything is implemented in JAX, including the environments. This allows for extremely fast training and evaluation of agents, because the entire loop of environment simulation, agent interaction, and policy updates can be compiled as a single XLA program and executed on CPUs, GPUs, or TPUs. More specifically, the implementations in Dopamax follow the Anakin Podracer architecture -- see this paper for more details.

Supported Algorithms

• Proximal Policy Optimization (PPO)
• Deep Q-Network (DQN)
• Deep Deterministic Policy Gradients (DDPG)
• Twin Delayed DDPG (TD3)
• Soft Actor Critic
• AlphaZero

Installation

Dopamax can be installed with:

pip install dopamax

This will install the dopamax Python package, as well as a command-line interface (CLI) for training and evaluation. Note that only the CPU version of JAX is installed by default. If you would like to use a GPU or TPU, you will need to install the appropriate version of JAX. See the JAX installation instructions.

Usage

After installation, the Dopamax CLI can be used to train and evaluate agents:

dopamax --help

Dopamax uses Weights and Biases (W&B) for logging and artifact management. Before using the CLI for training and evaluation, you must first make sure you have a W&B account (it's free) and have authenticated with wandb login.

Training

Agent's can be trained using the dopamax train command, to which you must provide a configuration file. The configuration file is a YAML file that specifies the agent, environment, and training hyperparameters. You can find examples in the examples directory. For example, to train a PPO agent on the CartPole environment, you would run:

dopamax train --config examples/ppo-cartpole/config.yaml

Note that all of the example config files have a random seed specified, so you will get the same result every time you run the command. The seeds provided in the examples are known to result in a successful run (with the given hyperparameters). To get different results on each run, you can remove the seed from the config file.

Evaluation

Once you have trained some agents, you can evaluate them using the dopamax evaluate command. This will allow you to specify a W&B agent artifact that you'd like to evaluate (these artifacts are produced by the training runs and contain the agent hyperparameters and weights from the end of training). For example, to evaluate a PPO agent trained on CartPole, you might use a command like:

dopamax evaluate --agent_artifact CartPole-PPO-agent:v0 --num_episodes 100

where --num_episodes 100 signals that you would like to rollout the agent's policy for 100 episodes. The minimum, mean, and maximum episode reward will be logged back to W&B. If you would additionally like to render the episodes and have then logged back to W&B, you can provide the --render flag. But note that this will usually significantly slow down the evaluation process since environment rendering is not a pure JAX function and requires callbacks to the host. You should usually only use the --render flag with a small number of episodes.

See Also

Some of the JAX-native packages that Dopamax relies on:

• sotetsuk/pgx
• deepmind/mctx
• deepmind/rlax
• google/brax