jovsa / evolution-strategies-exploration

Contains implementation of: Tim Salimans Et al. “Evolution Strategies as a Scalable Alternative to Reinforcement Learning”. Arxiv.org. https://arxiv.org/pdf/1703.03864.pdf.

Home Page:https://medium.com/@jovansardinha/an-exploration-into-evolution-strategies-97c42122c486

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

An Exploration into Evolution Strategies

Accompanying write-up can be found on my medium post

objectives:
1 — Implement evolution strategies from scratch and use it to optimize the weights of a neural network on the task of MNIST digit recognition.
2 — Find a good set of hyperparameters of the algorithm that achieve the best results after 12 hours of training.
3 — Distribute the above across the cores of a computer (going to 4 cores). Analyze the speedup observed when going from 1 core to a 4 core implementation.

Folder structure:

.
├── evolution/
|   ├── tests/  # contains all test    
|   ├── __inti__.py									
|   ├── es.py	  # implimentation of ES class  
|   └── main.py  # main file to execute   
├── notebooks/
|   ├── Analyzing Results of Best Hyperparameter.ipynb  # analyzing best parameter results  
|   ├── Analyzing Results of Hyperparameter Search.ipynb # post analysis for hyperparameter tuning  
|   ├── Analyzing Times of Runs by Number of Workers.ipynb # post analysis of runtimes  
|   ├── MNIST - Keras.ipynb		 # MNIST keras models  
|   └── bare bones implementation of NES - karpathy.ipynb # karpathy ES starter  
|── .gitignore
|── .README.md
└── requirements.txt  # list of packages used

Built and tested on:
operating system: Ubuntu 16.04.2 LTS
python version: 3.5.2
pip version: 9.0.1

About

Contains implementation of: Tim Salimans Et al. “Evolution Strategies as a Scalable Alternative to Reinforcement Learning”. Arxiv.org. https://arxiv.org/pdf/1703.03864.pdf.

https://medium.com/@jovansardinha/an-exploration-into-evolution-strategies-97c42122c486


Languages

Language:Jupyter Notebook 97.1%Language:Python 2.9%