search

JuliaReinforcementLearning / ReinforcementLearning.jl

A reinforcement learning package for Julia
https://juliareinforcementlearning.org
Other
587 stars 112 forks source link

In DDPG: Add support for vector actions #138

Closed oysteinsolheim closed 3 years ago

oysteinsolheim commented 4 years ago

Currently only scalar actions are supported, how far up on the priority list is it to expand on this?

findmyway commented 4 years ago

I think it's relatively easy to support actions of a vector in DDPG, but the main issue here is that we lack environments with this kind of actions 😰

Do you know any environments that are good testbeds for DDPG with multiple actions? @jbrea

jbrea commented 4 years ago

Do you know any environments that are good testbeds for DDPG with multiple actions?

There is e.g. Mujoco, but I don't have a license. There are also some environments in Gym (GymEnv in RLEnvironments), e.g. I think most of the bullet envs. And I think it would also be cool to create some environments with RigidBodySims... But I don't know right now which environment would be the simplest to just run a couple of tests.

findmyway commented 4 years ago

I may improve existing algorithms to support vector actions in the near future. Ref JuliaReinforcementLearning/ReinforcementLearningZoo.jl#102

Action space We include tasks with both discrete and continuous action spaces, and of varying action dimension with up to 56 dimensions in the initial release of RL Unplugged.