[RL-baseline] Model v4, experiment #1 #39
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The policy network for model v4 for REINFORCE with Baseline is essentially the same network as in v2, but the actor and critic heads have an additional fully connected layer each similar to v3. This tweak was added with the hopes of seeing the initial gains in reward that we observed with model v3 but with the elevated sustained reward value that we observed in v2.
The action sets are the same as in Model v3. For this experiment, action set #0 is chosen:
[0.0, 0.0, 0.0], # no action
[0.0, 0.8, 0.0], # throttle
[0.0, 0.3, 0.0], # throttle
[0.0, 0.0, 0.6], # break
[0.0, 0.0, 0.2], # break
[-0.9, 0.0, 0.0], # left
[-0.5, 0.0, 0.0], # left
[-0.2, 0.0, 0.0], # left
[0.9, 0.0, 0.0], # right
[0.5, 0.0, 0.0], # right
[0.2, 0.0, 0.0], # right
The results weren't as good as expected. Around the episode 2.5k mark the network managed to get a Running Reward of 258 but it quickly dropped to subzero values. Around the episode 18k mark the Running Reward became positive again and with additional training it could eventually lead to better results, but we've managed better results in other experiments at 20k episodes. Final Running Reward is 65.
Sample video below:
https://user-images.githubusercontent.com/1465235/112986275-96099e00-9161-11eb-9b24-d74a435918ec.mp4