Benchmarking Evolutionary Reinforcement Learning (pronounced "barrel"... sort of)

A collaborative project for aggregating benchmarks of evolutionary algorithms on common reinforcement learning benchmarks, based on Cambrian.jl.
Contribution guidelines are available here.

Algorithms:

• NEAT
• CGP

Environments:

• Iris classification
• XOR
• Gym classic control
• Atari on RAM

Algorithms:

• HyperNEAT (2xfeedforward & recurrent ANNs)
• CMA-ES (2xfeedforward & recurrent ANNs)
• population-based REINFORCE
• AGRN
• TPG
• grammatical evolution

Environments:

• mujoco
• pybullet
• mario

Customizable fitness:

• sum of reward over an episode
• Novelty?
• MAP-Elites

CLI interaction

• Parseable arguments

Non-Cambrian algorithms

• Interaction through a simplified interface with the BERL environments

To run a selection of algorithms on BERL benchmarks, please:

1. Toggle the algorithms and environments you want in the YAML config files.
2. Run

run_berl()

To only run a pair of algorithm and environment, you can also use:

start_berl(algo_name::String, env_name::String; env_params...)

env_params are the specific game names (such as "CartPole-v1") when "atari" or "gym" are selected as env_name.