policy_old完全看不出作用 about ppo-pytorch HOT 6 OPEN

你这个流程完全等同于

1. policy 生成数据
2. policy 迭代更新
3. loop

from ppo-pytorch.

You are right, but the author is also right to do so.

I think the current process is not redundant:

1. Update the policy K(=80) times per epoch. If the KL divergence needs to be calculated inside the epoch, policy_old must be retained.
2. In fact, KL divergence can be used as a trick for PPO algorithm. The torch version PPO, implemented by OpenAI Spinning Up, early stops the epoch just when the KL divergence is too large.

from ppo-pytorch.

感谢回复，你的观点我认同，old Policy确实可以用于计算KL散度，让模型本次更新不至于过大

另外当我知道PPO绕了一大圈，最后跟我说另外个分布就是上次的分布时，我是真的想掀桌子

整个PPO完全就是工程化的时候，看到for循环迭代缺少个阻尼项，而做的尝试。实际写出的论文却如此晦涩

from ppo-pytorch.

Although the motivation of PPO might not be that simple, I guess partial of PPO's motivation were found when applying TRPO in engineering.

Actually, introducing KL divergence into PPO to early stop is just a trick. The authors of PPO paper not meant to do that.

In my opinion, PPO is motivated to solve TRPO's computational complexity problem. Instead of computing KL divergence (very slow) like TRPO, PPO(clip version) simply limit the policy's update with a clip() function.

You can check it deeper by viewing OpenAI Spinning Up for PPO, which I have cited it as an url below:)

PPO is motivated by the same question as TRPO: how can we take the biggest possible improvement step on a policy using the data we currently have, without stepping so far that we accidentally cause performance collapse? Where TRPO tries to solve this problem with a complex second-order method, PPO is a family of first-order methods that use a few other tricks to keep new policies close to old. PPO methods are significantly simpler to implement, and empirically seem to perform at least as well as TRPO.

There are two primary variants of PPO: PPO-Penalty and PPO-Clip.

PPO-Penalty approximately solves a KL-constrained update like TRPO, but penalizes the KL-divergence in the objective function instead of making it a hard constraint, and automatically adjusts the penalty coefficient over the course of training so that it’s scaled appropriately.

PPO-Clip doesn’t have a KL-divergence term in the objective and doesn’t have a constraint at all. Instead relies on specialized clipping in the objective function to remove incentives for the new policy to get far from the old policy.

——from OpenAI Spinning Up for PPO

from ppo-pytorch.

嗯，是的。如果从TRPO出发，PPO的改进是成功的

再次感谢你的回答，我后续会看下OpenAI的版本

from ppo-pytorch.

You are welcomed :)

from ppo-pytorch.