OpenAI ES
A variant of evolution strategies proposed by OpenAI that does not explicitly maintain distribution parameters, instead applying Gaussian perturbations to model parameters and updating them along the reward-weighted average direction.
Timeline (last 90 days)
The ES variant discussed in this paper is in the same lineage as OpenAI ES: rather than explicitly maintaining distribution parameters, it applies Gaussian perturbations to the current model parameters to obtain offspring and updates using the reward-weighted average direction as weights.
The OpenAI ES form is popular in reinforcement learning and black-box optimization because it is easy to parallelize, does not rely on backpropagation, and is somewhat robust to reward scale.
All Facts (2)
The ES variant discussed in this paper is in the same lineage as OpenAI ES: rather than explicitly maintaining distribution parameters, it applies Gaussian perturbations to the current model parameters to obtain offspring and updates using the reward-weighted average direction as weights.
50%UnverifiedThe OpenAI ES form is popular in reinforcement learning and black-box optimization because it is easy to parallelize, does not rely on backpropagation, and is somewhat robust to reward scale.
50%