ESTool

Implementation of various Evolution Strategies, such as GA, Population-based REINFORCE (Section 6 of Williams 1992), CMA-ES and OpenAI's ES using common interface.

CMA-ES is wrapping around pycma.

Notes

The tool last tested using the following configuration:

• NumPy 1.13.3 (1.14 has some annoying warning).

• OpenAI Gym 0.9.4 (breaks for 0.10.0+ since they changed the API).

• cma 2.2.0, basically 2+ should work.

• PyBullet 1.6.3 (possible that newer versions might work, but have not tested).

• Python 3, although 2 might work.

• mpi4py 2

Backround Reading:

A Visual Guide to Evolution Strategies

Evolving Stable Strategies

Using Evolution Strategies Library

To use es.py, please check out the simple_es_example.ipynb notebook.

The basic concept is:

solver = EvolutionStrategy()
while True:

  # ask the ES to give us a set of candidate solutions
  solutions = solver.ask()

  # create an array to hold the solutions.
  # solver.popsize = population size
  rewards = np.zeros(solver.popsize)

  # calculate the reward for each given solution
  # using your own evaluate() method
  for i in range(solver.popsize):
    rewards[i] = evaluate(solutions[i])

  # give rewards back to ES
  solver.tell(rewards)

  # get best parameter, reward from ES
  reward_vector = solver.result()

  if reward_vector[1] > MY_REQUIRED_REWARD:
    break

Parallel Processing Training with MPI

Please read Evolving Stable Strategies article for more demos and use cases.

To use the training tool (relies on MPI):

python train.py bullet_racecar -n 8 -t 4

will launch training jobs with 32 workers (using 8 MPI processes). the best model will be saved as a .json file in log/. This model should train in a few minutes on a 2014 MacBook Pro.

If you have more compute and have access to a 64-core CPU machine, I recommend:

python train.py name_of_environment -e 16 -n 64 -t 4

This will calculate fitness values based on an average of 16 random runs, on 256 workers (64 MPI processes x 4). In my experience this works reasonably well for most tasks inside config.py.

After training, to run pre-trained models:

python model.py bullet_ant log/name_of_your_json_file.json

Self-Contained Cartpole Swingup Task

If you don't want to install a physics engine, try it on the cartpole_swingup task that doesn't have any dependencies:

Training command:

python train.py cartpole_swingup -n 8 -e 1 -t 4 --sigma_init 1.0

After 400 generations, the final average score (over 32 trials) should be over 900. You can run it with this command:

python model.py cartpole_swingup log/cartpole_swingup.cma.1.32.best.json

If you haven't bothered to run the previous training command, you can load the pre-trained version:

python model.py cartpole_swingup zoo/cartpole_swingup.cma.json

Self-Contained Slime Volleyball Gym Environment

Here is an example for training slime volleyball gym environment:

Training command:

python train.py slimevolley -n 8 -e 8 -t 4 --sigma_init 0.5

Pre-trained model:

python model.py slimevolley zoo/slimevolley.cma.64.96.best.json

PyBullet Envs

Another example: to run a minitaur duck model, run this locally:

python model.py bullet_minitaur_duck zoo/bullet_minitaur_duck.cma.256.json

In the .hist.json file, and on the screen output, we track the progress of training. The ordering of fields are:

• generation count
• time (seconds) taken so far
• average fitness
• worst fitness
• best fitness
• average standard deviation of params
• average timesteps taken
• max timesteps taken

Using plot_training_progress.ipynb in an IPython notebook, you can plot the traning logs for the .hist.json files. For example, in the bullet_ant task:

You need to install mpi4py, pybullet, gym etc to use various environments. Also roboschool/Box2D for some of the OpenAI gym envs.

On Windows, it is easiest to install mpi4py as follows:

• Download and install mpi_x64.Msi from the HPC Pack 2012 MS-MPI Redistributable Package
• Install a recent Visual Studio version with C++ compiler
• Open a command prompt

git clone https://github.com/mpi4py/mpi4py
cd mpi4py
python setup.py install

Modify the train.py script and replace mpirun with mpiexec and -np with -n

Citation

If you find this work useful, please cite it as:

@article{ha2017evolving,
  title   = "Evolving Stable Strategies",
  author  = "Ha, David",
  journal = "blog.otoro.net",
  year    = "2017",
  url     = "http://blog.otoro.net/2017/11/12/evolving-stable-strategies/"
}