Evolving-Reservoirs-for-Meta-Reinforcement-Learning

Code for the Evolving-Reservoirs-for-Meta-Reinforcement-Learning (ER-MRL) paper, presented at the Evostar 2024 conference [Long Talk]. Our goal is to study the following question : How neural structures, optimized at an evolutionary scale, can enhance the capabilities of agents to learn complex tasks at a developmental scale?

To achieve this, we adopt a computational framework based on meta reinforcement learning, modeling the interplay between evolution and development. At the evolutionary scale, we evolve reservoirs, a family of recurrent neural networks generated from hyperparameters. These evolved reservoirs are then utilized to facilitate the learning of a behavioral policy through reinforcement learning. This is done by encoding the environment state through the reservoir before providing it to the agent's policy. We refer to these agents, integrating a reservoir and a policy network, as ER-MRL agents. Our repository provides:

Python scripts to :

evolve ER-MRL agents with one or multiple reservoir(s) on the same environment
evolve ER-MRL agents with multiple reservoir on different environments to study generalization
evaluate evolved ER-MRL agents against standard RL agents with the different setups mentioned above
create your own reservoir architecture inside an ER-MRL agent or mask observations from an environment

A tutorial to parallelize our method :

Jupyter notebooks to :

Installation

1- Get the repository

git clone https://github.com/corentinlger/ER-MRL.git
cd ER-MRL/

2- Install the dependencies

python -m venv myvenv
source myvenv/bin/activate
pip install -r requirements.txt
pip install -e .

Usage

Evolve reservoir of ER-MRL agents on the same environment

To evolve and find the best reservoir structure within an ER-MRL agent on a specific task, you can use scripts/evolve_res.py. See the complete list of parameters (environment choice, number of training timesteps ...) you can use while running the evolution in this script.

python3 scripts/evolve_res.py --env_id HalfCheetah-v4 --h_test test_experiment --training_steps 300000 --nb_trials 100 --sampler Cmaes

We recommend runing these evolution phases on a remote cluster because they can rapidly become computanionally expensive. To do so, you can follow the tutorials present in the parallelization tutorials directory.

Test ER-MRL agents equipped with the best evolved reservoir

If you want to test the best evolved ER-MRL agent against standard RL agents, you can use the following command (make sure you provide the parameters corresponding to the ones used in the evolution phase):

python3 scripts/test.py --env_id HalfCheetah-v4 --h_test test_experiment --HP_training_steps 300000

Analyze the results

To analyze the results obtained both during the evolution and the testing phases, you can follow the the steps described in this jupyter notebook.

Use a multi-reservoirs setup

If you want to evolve agents containing multiple reservoirs instead of one, use the scripts/evolve_multi_res.py to run the evolution, and scripts/test_multi_res.py to test the evolved agents performance (you will need to specify the number of reservoirs desired).

Study generalization of neural structures by evolving reservoirs of ER-MRL agents on different environments

To evolve the reservoirs of ER-MRL agents on a diversity of environments and test them on unseen ones during the evolution phase, you can run the following command:

python3 scripts/evolve_generalization.py --nb_res 2 --env_type Ant_Swimmer --h_test generalization_test_experiment --training_timesteps 300000 --nb_trials 100 --sampler Cmaes

You can either use a predefined set of evolution environments (the one above includes Ant-v4 and Swimmer-v4), or go in this file and add a custom env_type to the generate_env_ids function.

Test ER-MRL agents with evolved reservoirs on new unseen tasks

And then test the evolved ER-MRL agents on a new unseen environment (HalfCheetah-v4 in this case) as follows :

python3 scripts/test_generalization.py --nb_res 2 --HP_env_type Ant_Swimmer --env_id HalfCheetah-v4 --h_test generalization_test_experiment

Citing

@inproceedings{leger2024evolving,
  title={Evolving Reservoirs for Meta Reinforcement Learning},
  author={L{\'e}ger, Corentin and Hamon, Gautier and Nisioti, Eleni and Hinaut, Xavier and Moulin-Frier, Cl{\'e}ment},
  booktitle={International Conference on the Applications of Evolutionary Computation (Part of EvoStar)},
  pages={36--60},
  year={2024},
  organization={Springer}
}