CHILD POET

CHILD POET is a modification to Enchanced POET designed to reduce resource the computational cost of the algorithm.

Simple glossary and definitions

• Agent model - refers to a neural net controlling the bipedal walker. Sometimes referred to simply as "model"
• Thetas - the neural net parameters of an agent model
• Niche - Refers to a particular implementation of an environment, with its configuration and cppn
• Environment - Refers to either only the Bipedal Walker Environment, or is synonymous with niche
• env_config - Configurations for a niche, including its name. Does not contain cppn.
• env_parameters - Parameters for a niche encoding, mostly everything that has to do with CPPN

POET and Enhanced POET

This repo contains implementation of the POET and Enhanced POET algorithms described in:

Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions

Enhanced POET: Open-Ended Reinforcement Learning through Unbounded Invention of Learning Challenges and their Solutions

An article on Uber Engineering Blog describing POET can be found here.

Requirements

• Fiber
• NEAT-Python
• OpenAI Gym

Run Enhanced POET locally

To run locally on a multicore machine

./run_poet_local.sh final_test

Run Enhanced POET on a Kubernetes cluster

Follow instructions here to create a persistent volume.

Then run the following command:

./run_poet_remote.sh final_test

To get the training logs:

fiber cp nfs:/persistent/logs/final_test . fiber cp nfs:/persistent/logs/poet_final_test poet_final_test

Run Enhanced POET on a computer cluster

To containerize and run the code on a computer cluster (e.g., Google Kubernetes Engine on Google Cloud), please refer to Fiber Documentation.

Run original POET

Use this legacy branch