Example scripts

The scripts provided here demonstrate key features of EvoTorch for new users and serve as starting points for solving custom problems.

Black-box optimization

bbo_vectorized.py: demonstrates single objective black-box optimization using a distribution-based algorithm, accelerated using vectorization on a single GPU/CPU.
moo_parallel.py: demonstrates multi-objective optimization using parallelization on all CPU cores without vectorization.

Reinforcement Learning

rl_gym.py: demonstrates how to solve a simple Gymnasium problem using the PGPE algorithm and ClipUp optimizer.

Paper re-implementation (RL)

The script rl_clipup.py reproduces experiments from the following paper:

Nihat Engin Toklu, Paweł Liskowski, and Rupesh Kumar Srivastava.
"ClipUp: a simple and powerful optimizer for distribution-based policy evolution."
International Conference on Parallel Problem Solving from Nature. Springer, Cham, 2020.
https://arxiv.org/abs/2008.02387

It allows you to train policies for the Lunar Lander, Walker-2D and Humanoid environments from Gymnasium as well as the PyBullet Humanoid.

Requirements

# Necessary: Used for configs and logging results to files or databases
pip install sacred
# Optional: box2d for Lunar Lander, mujoco for Walker-2D and Humanoid
pip install 'gymnasium[box2d,mujoco]'
# Optional: For PyBullet Humanoid
pip install pybullet 'gym<0.26'

Running experiments

The script rl_clipup.py is a command line utility for training and saving policies for classical gym environments. Its basic usage is as follows:

python rl_clipup.py -F RESULTS_DIR           \
    with env_name=GYM_ENVIRONMENT_ID  \
         policy=POLICY                \
         save_interval=SAVE_INTERVAL  \
         hyperparameter1=value1       \
         hyperparameter2=value2       \
         ...

where:

RESULTS_DIR is the directory into which the logs and the artifacts (pickle files containing the evolved policies and metadata) will be saved.
GYM_ENVIRONMENT_ID is a string such as "LunarLander-v2" or "Humanoid-v4", etc., indicating which environment is to be solved.
POLICY is a string which determines the architecture of the neural network that will be trained. By default, this is set as a single-hidden-layered network (the hidden layer having 16 neurons), represented by the string: "Linear(obs_length, 16) >> Tanh() >> Linear(16, act_length)" (where obs_length represents the length of the observation vector, and act_length represents the length of the action vector). To use a linear policy with bias, one can set this string as: "Linear(obs_length, act_length)". Alternatively, for having a linear policy without bias, one can set this string as: "Linear(obs_length, act_length, bias=False)".
SAVE_INTERVAL can be an integer n, which means that at every n generations, the current state of the evolved policy will be saved to disk. SAVE_INTERVAL can also be set as "last", which means that the evolved policy will be saved only at the end of the entire run. For large experiments, it is recommended to set this as an integer such as 10, or 50, etc.

To see the full list of available hyperparameters with their explanations and their default values, one can use the following shell command:

python rl_clipup.py print_config

One can also see the effects of changing hyperparameter(s), without executing the experiment, via:

python rl_clipup.py print_config with hyperparameter1=newvalue1 hyperparameter2=newvalue2 ...

Available pre-configurations

rl_clipup.py comes with pre-configurations for certain environments.

For example, the following shell command:

python rl_clipup.py -F RESULTS_DIR with lunarlander

...evolves a LunarLanderContinuous-v2 and saves it into RESULTS_DIR.

Also, the following command:

python rl_clipup.py -F RESULTS_DIR with pybullet_humanoid save_interval=50

...starts an evolutionary computation run for solving wrapped_humanoid_bullet:WrappedHumanoidBulletEnv-v0 (which is a wrapper around the classical gym environment pybullet_envs:HumanoidBulletEnv-v0 to make the environment available to gymnasium), and saves the policy at every 50 generations. Such an explicit save_interval value is recommended for pybullet humanoid, since the computational experiments for this environment last long, and one might want to look at how the current agent is behaving without having to wait until the end of the run.

Other available pre-configurations are walker (for the MuJoCo environment Walker-v4) and humanoid (for the MuJoCo environment Humanoid-v4).

Script for testing and/or visualizing policies: `rl_enjoy.py`

The policies trained/evolved by rl_clipup.py are saved as pickle files. These pickle files can then be tested with the help of rl_enjoy.py. The simplest usage is:

python rl_enjoy.py PICKLE_FILE

...which loads the policy saved in the specified pickle file, runs and visualizes it in an episode of its environment, and prints the obtained total reward.

For further help, one might use the following shell command:

python rl_enjoy.py --help