couple of improvements to MPO

[HenriDeh]

This simply

• modifies slightly the math of vec_to_tril to create more stable covariance matrices
• make the upperbound of the dual adaptive to let it be higher when rewards (and thus Q-values) have higher magnitudes.

[codecov]

Codecov Report

Merging #919 (e5f102f) into main (b54a0b0) will increase coverage by 0.00%.
The diff coverage is 93.33%.

@@           Coverage Diff           @@
##             main     #919   +/-   ##
=======================================
  Coverage   24.27%   24.28%           
=======================================
  Files         221      221           
  Lines        7739     7741    +2     
=======================================
+ Hits         1879     1880    +1     
- Misses       5860     5861    +1     

Impacted Files	Coverage Δ
...c/ReinforcementLearningCore/test/utils/networks.jl	46.42% <ø> (ø)
...tLearningZoo/src/algorithms/policy_gradient/mpo.jl	0.00% <0.00%> (ø)
...rc/ReinforcementLearningCore/src/utils/networks.jl	75.39% <100.00%> (+0.39%)	⬆️

... and 1 file with indirect coverage changes

[HenriDeh]

@jeremiahpslewis can you validate please ?

[jeremiahpslewis]

@HenriDeh I looked at this PR and did my best to review it based on things I felt I can contribute. In terms of computational correctness, I don't really have a clue for these functions, but I would suggest that you add unit tests for the functions (if only so that if future changes to the code lead to different results, we can investigate and learn whether we've broken something or it was broken all along...

[jeremiahpslewis]

see above

[HenriDeh]

All tests are passing on my machine. The failing test in e004bef seems to be a bug (a timer timed 4x10^233), quite impossible. I couldn't reproduce on my machine and it is unrelated to the changes of the PR. Hopefully this does not happen again in the CI job.

[jeremiahpslewis]

Looks awesome!