Hi!
I ran this project in Ubuntu20.04 successfully by using TASK="halfcheetah-medium-v2"; seed=0; python3 -u train_behavior.py --expid ${TASK}-baseline-seed${seed} --env $TASK --seed ${seed} and TASK="halfcheetah-medium-v2"; seed=0; python3 -u train_critic.py --expid ${TASK}-baseline-seed${seed} --env $TASK --seed ${seed}

Thus,I think your system is Ubuntu.

But when I ran these code in Windows10, some errors have happened.So I have a suggestion about README.md.

If someone want use Windows to run this project,he may use this code $env:TASK="halfcheetah-medium-v2"; $env:seed=0; python -u train_behavior.py --expid $env:TASK-baseline-seed$env:seed --env $env:TASK --seed $env:seed and $env:TASK="halfcheetah-medium-v2"; $env:seed=0; python -u train_critic.py --expid $env:TASK-baseline-seed$env:seed --env $env:TASK --seed $env:seed

Similarly,if he wants to train the policy,just rewrite the code as mentioned above.

Finally，I want to know,when I run this code successfully,the terminal show me that

and WandB has a figure

The progress bar for training behavior has consistently remained at 0%, suggesting that the behavior is not being trained. Is this a normal occurrence?

Hi!bro:
As a green hand,I want to ask a problem.Could you tell me how to do it?
When I use this code$ TASK="halfcheetah-medium-v2"; seed=0; python3 -u train_behavior.py --expid ${TASK}-baseline-seed${seed} --env $TASK --seed ${seed},terminal tells me that `$ : 无法将“$”项识别为 cmdlet、函数、脚本文件或可运行程序的名称。请检查名称的拼写，如果包括路径，请确保路径正确，然后再试一次。
所在位置 行:1 字符: 1
所在位置 行:1 字符: 33

• $ TASK="halfcheetah-medium-v2"; seed=0; python3 -u train_behavior.py
  • CategoryInfo : ObjectNotFound: (seed=0:String) [], CommandNotFoundException
  • FullyQualifiedErrorId : CommandNotFoundException`

I think that this error may be caused by system.My system is Windows10,and the environment is python3.7.
Your system may be Ubuntu20.04?

Hi!

I was wondering the following, the Dirac policy only returns actions in the range of -1 and 1, since the activation function in the last layer is the tanh function. Thus, I think for environments that have an action space, that is outside of that range, your method might under perform, or am I missing something?

Maybe it would be convenient to scale the policy output by the action space of the environment.