Skip to content

Latest commit

 

History

History
31 lines (19 loc) · 1.28 KB

README.md

File metadata and controls

31 lines (19 loc) · 1.28 KB
Raw

WorldModels

An implementation of the ideas from this paper https://arxiv.org/pdf/1803.10122.pdf

Code base adapted from https://github.com/hardmaru/estool

This repo. was created with an objective to learn Generative modeling in an enclosed Open-AI's environment with Evolutionary Strategy using the CMA-ES algorithm. There are some ongoing experiments for the next steps with different backbones and Attention units.

For full installation and run instructions:

  • Create the anaconda environment file from requirement.txt
  • Base frameworks in use are Keras-GPU(v2.3.1) and Tensorflow-GPU(v2.1)
  • Run the notebooks for analysis of the experiments based on the weights and collected data
  • For rendering the benchmarks, in OpenAI Gym environment:
python model.py car_racing --filename ./controller/car_racing.cma.xxxxx.best.json --render --final_mode
  • The videos for some agents reaching some good scores are here
  • The complete instructions for rendering and benchmark will be updated soon...

Architecture for the VAE (Backbone)

VAE

Architecture for the RNN

RNN

Box/Machine configuration: Experimentations done on a AMD's 8-core CPU with a Nvidia's GeForce RTX-2070 (Turing)