Natural Evolution Strategies

Implementation from the paper (Evolution Strategies as a Scalable Alternative to Reinforcement Learning)[https://arxiv.org/pdf/1703.03864.pdf]

Note: This is under development

Getting Started:

• Create a virtual environment

  mkvirtualenv -p python3 <env-name>
  

• Install the dependencies

  pip install -r requirements.txt
  

• Run the algorithm

  python src/run.py
  

• Check help for a list of parameters

  python src/run.py --help
  

• Change the objective function and initial solution accordingly to suit your problem statement

  def fitness(solution):
      """Objective function for optimization"""
      # this is the objective that NES will try to maximize
      objective = ...
      return objective
  .
  .
  .
  
  # this is where you define your solution space
  init_solution = np.random.random(size=3)
  

TODO:

• Implement a basic version
• Try the basic solution on a toy problem
• Implement a distributed version
• Try the distributed version on a toy problem
• Try the distributed version on larger problem