This PR adds strict typing to the output of `update` and `learn` in all
policies. This will likely be the last large refactoring PR before the
next release (0.6.0, not 1.0.0), so it requires some attention. Several
difficulties were encountered on the path to that goal:
1. The policy hierarchy is actually "broken" in the sense that the keys
of dicts that were output by `learn` did not follow the same enhancement
(inheritance) pattern as the policies. This is a real problem and should
be addressed in the near future. Generally, several aspects of the
policy design and hierarchy might deserve a dedicated discussion.
2. Each policy needs to be generic in the stats return type, because one
might want to extend it at some point and then also extend the stats.
Even within the source code base this pattern is necessary in many
places.
3. The interaction between learn and update is a bit quirky, we
currently handle it by having update modify special field inside
TrainingStats, whereas all other fields are handled by learn.
4. The IQM module is a policy wrapper and required a
TrainingStatsWrapper. The latter relies on a bunch of black magic.
They were addressed by:
1. Live with the broken hierarchy, which is now made visible by bounds
in generics. We use type: ignore where appropriate.
2. Make all policies generic with bounds following the policy
inheritance hierarchy (which is incorrect, see above). We experimented a
bit with nested TrainingStats classes, but that seemed to add more
complexity and be harder to understand. Unfortunately, mypy thinks that
the code below is wrong, wherefore we have to add `type: ignore` to the
return of each `learn`
```python
T = TypeVar("T", bound=int)
def f() -> T:
return 3
```
3. See above
4. Write representative tests for the `TrainingStatsWrapper`. Still, the
black magic might cause nasty surprises down the line (I am not proud of
it)...
Closes#933
---------
Co-authored-by: Maximilian Huettenrauch <m.huettenrauch@appliedai.de>
Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
Our default choice: remove the done flag penalty, will soon converge to ~280 reward within 100 epochs (10M env steps, 3~4 hours, see the image below)
If the done penalty is not removed, it converges much slower than before, about 200 epochs (20M env steps) to reach the same performance (~200 reward)
BipedalWalker-BDQ
To demonstrate the cpabilities of the BDQ to scale up to big discrete action spaces, we run it on a discretized version of the BipedalWalker-v3 environment, where the number of possible actions in each dimension is 25, for a total of 25^4 = 390 625 possible actions. A usaual DQN architecture would use 25^4 output neurons for the Q-network, thus scaling exponentially with the number of action space dimensions, while the Branching architecture scales linearly and uses only 25*4 output neurons.
522f7fbf98
