Utilizing Evolution Strategies to Train Transformers in Reinforcement Learning

Matyáš Lorenc, Roman Neruda
1/23/2025
cs.LGcs.NE

No official code found. Best implementation: Mafi412/Evolution-Strategies-and-Decision-Transformers (3 stars, MIT, last commit 2025-05-05).

Abstract

We explore the capability of evolution strategies to train an agent with a policy based on a transformer architecture in a reinforcement learning setting. We performed experiments using OpenAI's highly parallelizable evolution strategy to train Decision Transformer in the MuJoCo Humanoid locomotion environment and in the environment of Atari games, testing the ability of this black-box optimization technique to train even such relatively large and complicated models (compared to those previously tested in the literature). The examined evolution strategy proved to be, in general, capable of achieving strong results and managed to produce high-performing agents, showcasing evolution's ability to tackle the training of even such complex models.

View on arXivView PDF

Code Implementations(1)

Mafi412/Evolution-Strategies-and-Decision-TransformersCommunity100%
MIT

This is a repository containing codebase and experimental data for paper Utilizing Evolution Strategies to Train Transformers in Reinforcement Learning.

33PythonJan 22, 20251 years agoMIT

Similar Papers

Transformers in Reinforcement Learning: A Survey

86% match1 repo542023

Solving Deep Reinforcement Learning Tasks with Evolution Strategies and Linear Policy Networks

84% match1 repo52024

Supervised Pretraining Can Learn In-Context Reinforcement Learning

83% match1 repo792023

Transformers as Decision Makers: Provable In-Context Reinforcement Learning via Supervised Pretraining

83% match1 repo102026

Transformers as Policies for Variable Action Environments

83% match1 repo92023

A Survey on Transformers in Reinforcement Learning

82% match2023

Cite this paper

@article{lorenc2025utilizing,
  title  = {Utilizing Evolution Strategies to Train Transformers in Reinforcement Learning},
  author = {Matyáš Lorenc and Roman Neruda},
  year   = {2025},
  eprint = {2501.13883},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2501.13883v2}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2501.13883.svg)](https://codeofpaper.com/paper/2501.13883)

Discussion