Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Is max timestep limit useful? #158

Closed
HeegerGao opened this issue Jul 15, 2020 · 1 comment
Closed

Is max timestep limit useful? #158

HeegerGao opened this issue Jul 15, 2020 · 1 comment

Comments

@HeegerGao
Copy link

Hi, I am new to metaworld. I find that you set a max_path_length in metaworld/envs/mujoco/mujoco_env.py. I find that you mentioned this in #39, where you said it is because Mujoco environments have a max_path_length of 150.

But I find that in other mujoco envs (like openai gym), there is no such timestep limit. So I just delete line 101 and 102 in metaworld/envs/mujoco/mujoco_env.py to unrestrict the limit. Then I just run an env from your benchmark envs for 1000 timesteps and I don't encounter any problem. So I just wonder if this timestep limit is useful.

@ryanjulian
Copy link
Contributor

Reposting my response from #157

My short answer is "yes," and you can read below for why it exists.

@HeegerGao thanks for bringing up this feedback!

The max_path_length limit is intended to ensure that results generated with metaworld are reproducible across research papers. The performance of two algorithms can be drastically different using two different max_path_length numbers, given the same number of environment samples. This is especially true in the case of meta-learning, which typically uses trajectories during the adaptation step. It also functions to give users a reasonable idea of how long an environment takes to solve.

I'll take this issue as feedback for future revisions of the benchmark. It may be artificially-constraining to limit the max_path_length, when we can already measure the total environment steps taken during training regardless of max_path_length (and in the case of meta-RL, the total environment steps prior to adaptation).

If you plan on writing a research paper, please note: regardless of your perspective on design decisions, it's never valid to report results in a research paper as "metaworld" using a modified version of the benchmark code. If a limitation is particularly-cumbersome and you need to change it, we ask that you present the results from both the modified and unmodified version of the benchmark in all figures/tables/plots, and make clear in the main text the exact modifications you made.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants