[RLlib; docs] RLlib docs redo: New API stack Episodes (SingleAgentEpisode). #46985
Conversation
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
There was a problem hiding this comment.
LGTM. We are moving forward here. Very nice docs for the complex episodes. Amazing how simpel and clear this is described @sven1977!!
|
|
||
| # Get the very first observation ("reset observation"). Note that a single observation | ||
| # is returned here (not a list of size 1 or a batch of size 1). | ||
| check(episode.get_observations(0), "obs_0") |
There was a problem hiding this comment.
Very nice!! With this we can be sure its the observation at timestep 0.
There was a problem hiding this comment.
Awesome diagrams. Makes the lookback buffer cristal clear!
Only thing: Describe the length of the lookback in this case :)
There was a problem hiding this comment.
Really nice! I think this is well understandable.
| :py:class:`~ray.rllib.env.single_agent_episode.SingleAgentEpisode` for single-agent setups | ||
| and :py:class:`~ray.rllib.env.multi_agent_episode.MultiAgentEpisode` for multi-agent setups. | ||
| The data is translated from this `Episode` format to tensor batches (including a possible move to the GPU) | ||
| only immediately before a neural network forward pass. |
There was a problem hiding this comment.
Maybe we add here that this happens inside of the connector pipeline?
| :end-before: rllib-sa-episode-03-end | ||
|
|
||
|
|
||
| To illustrate the differences between the data stored in a non-finalized episode vs the same data stored in |
|
|
||
| To illustrate the differences between the data stored in a non-finalized episode vs the same data stored in | ||
| a finalized one, take a look at this complex observation example here, showing the exact same observation data in two | ||
| episodes (one not finalized the other finalized): |
There was a problem hiding this comment.
not finalized -> non-finalized?
|
|
||
| **Complex observations in a finalized episode**: The entire observation record is a single (complex) dict matching the | ||
| gym environment's observation space. At the leafs of the structure are `np.NDArrays` holding the individual values of the leaf. | ||
| Note that these `NDArrays` have an extra batch dim (axis=0), whose length matches the length of the episode stored (here 3). |
There was a problem hiding this comment.
Either both np.NDArray or both NDArray I'd say. What do you think?
|
|
||
| Another useful getter argument (besides `fill`) is the `neg_index_as_lookback` boolean argument. | ||
| If set to True, negative indices are not interpreted as "from the end", but as | ||
| "into the lookback buffer". This allows you to loop over a range of global timesteps |
There was a problem hiding this comment.
This is a tiny tiny bit ambiguous when the "lookback" is actually not into the "lookback buffer" but just backwards from the considered global timestep. So, precisely it would be "from the global timestep (if needed into the lookback buffer)", wouldn't it?
There was a problem hiding this comment.
I think this is fine.
Note that we do not support indexing by global (actual gym episode) timestep yet. It's probably a usecase that's never really needed.
So if you have an episode chunk from global timestep 100 to global timestep 200, as a user, you would normally just loop through it (from local_index 0 to local_index 100) and then - at each local_index - say: I want to do 4x-framestacking from local_index - 3 to local_index + 1. This means that if local_index is 0 or 1 or 2, you'd have a negative start index and for these few cases you want to tell the getter that these negative indices indeed mean "go into the lookback buffer" instead of "count from the end".
In other words: If you want to index into the lookback buffer, you have to provide a negative number, b/c index=0 is always the first item after(!) the lookback buffer.
| # range. Set to 0 (beginning of lookback buffer) if result is a negative | ||
| # index. | ||
| if neg_indices_left_of_zero: | ||
| if neg_index_as_lookback: |
There was a problem hiding this comment.
Nice change! This is smoother and sounds more correct :)
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
Signed-off-by: sven1977 <[email protected]>
…sode). (ray-project#46985) Signed-off-by: Dev <[email protected]>
RLlib docs redo: New API stack Episodes (SingleAgentEpisode).
Why are these changes needed?
Related issue number
Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.method in Tune, I've added it in
doc/source/tune/api/under thecorresponding
.rstfile.