codecov[bot]
commented
4 years ago Codecov Report
:exclamation: No coverage uploaded for pull request base (
master@5b76ea7). Click here to learn what that means. The diff coverage is83.06%.
@@ Coverage Diff @@
## master #13 +/- ##
=========================================
Coverage ? 76.02%
=========================================
Files ? 45
Lines ? 2886
Branches ? 0
=========================================
Hits ? 2194
Misses ? 692
Partials ? 0| Impacted Files | Coverage Δ | |
|---|---|---|
| src/evaluating_rewards/envs/mujoco.py | 98.16% <ø> (ø) |
|
| src/evaluating_rewards/analysis/stylesheets.py | 71.42% <ø> (ø) |
|
| ...uating_rewards/analysis/plot_divergence_heatmap.py | 66.66% <0%> (ø) |
|
| src/evaluating_rewards/experiments/comparisons.py | 0% <0%> (ø) |
|
| src/evaluating_rewards/scripts/train_regress.py | 37.5% <0%> (ø) |
|
| ...luating_rewards/experiments/point_mass_analysis.py | 78.12% <0%> (ø) |
|
| tests/test_scripts.py | 100% <100%> (ø) |
|
| ...c/evaluating_rewards/analysis/gridworld_heatmap.py | 96.22% <100%> (ø) |
|
| tests/test_comparisons.py | 100% <100%> (ø) |
|
| src/evaluating_rewards/experiments/synthetic.py | 88.78% <100%> (ø) |
|
| ... and 8 more |
Continue to review full report at Codecov.
Legend - Click here to learn more
Δ = absolute <relative> (impact),ø = not affected,? = missing dataPowered by Codecov. Last update 5b76ea7...ea0fcf8. Read the comment docs.
AdamGleave
Uses a more principled initialization of affine parameters to minimize least square error, rather than previous heuristic method matching mean and s.d. (which handles opposite rewards badly: sets scale to identity when should be zero).
Also add support for optimization via alternating minimization of affine parameters (analytic, closed form) and potential shaping weights (gradient descent). This worked well in the tabular setting. Seems to neither help nor hurt in the function approximator setting. Disabled by default.