[Feature] Add Trust-region algorithm searcher

[yhna940]

No description provided.

[yhna940]

We chose this algorithm because it showed the best performance in bbo benchmark experiments.
It differs from the original implementation in that it supports max mode and discrete variables.

[Ref]
https://arxiv.org/pdf/2104.10201v2.pdf
https://arxiv.org/abs/1910.01739
https://github.com/uber-research/TuRBO

[github-actions]

Coverage report

The coverage rate is 85.17%

The branch rate is 66.75999999999999%

81% of new lines are covered.

Diff Coverage details (click to unfold)

mmtune/ray/searchers/init.py

100.0% of new lines are covered

mmtune/mm/tasks/disc_test_func.py

97.14285714285714% of new lines are covered

mmtune/mm/tasks/init.py

100.0% of new lines are covered

mmtune/mm/tasks/cont_test_func.py

94.24083769633508% of new lines are covered

mmtune/ray/searchers/trust_region.py

68.5344827586207% of new lines are covered

[nijkah]

How about using args = self.args as in other tasks?

[nijkah]

How about using args = self.args as in other tasks?

[nijkah]

Docstring is required.

[nijkah]

@yhna940
The coverage rate seems much low. More unit tests should be required.
Please let me know if you need any help filling it.

[nijkah]

Should the previous class Sphere be deprecated?

[yhna940]

Since the sphere function alone does not seem to be a sufficient optimization test, a new Test function including it has been added. I used the nevergrad benchmark (noise free version).