Add tests

[ptoman]

Adds a variety of tests to the GridInterpolations.jl framework for simplex and multilinear interpolation.

===

This supersedes a previous pull request that included interpolations using k-nearest neighbors, re: a discussion between @zsunberg and I. This PR does not include kNN material. We convinced ourselves to exclude kNN because it's too much work for too little/no gain. Specifically:

• kNN isn't really a form of grid interpolation, and it's odd to have an interface that feeds in a set of points. If we're going to do incorporate this material, we really should instead have an interface-or-Julia-equivalent for "interpolation", of which grid interpolation and kNN interpolation are both subtypes. This, though, may start to border on overengineering.... (Side-note that kNN on random points might perform much better than kNN on a grid, if Bergstra and Bengio's 2012 findings and explanation that some dimensions matter more than others for finding good values holds in spaces that have high dimensionality but low effective dimensionality -- which seems like it might hold for many POMDP-useful-style spaces. This is another reason to think that kNN should be implemented independently from grids.)
• kNN-on-a-grid was found to perform quite poorly for interpolation tasks in the same evaluation from December 2016. It was slower than simplex and multilinear interpolation, and it was more inaccurate than either of them as well. Since it is fully dominated, we can't in good conscience recommend others use it / make it available for others to use.
• Maintenance has the potential to be challenging. The kNN implementations more than doubled the dependencies for GridInterpolations.jl, which requires staying on top of their releases for bugs and hoping that the API stays consistent, and extra headaches if it doesn't. Additionally, we have the challenge of the more code, the more difficulty in spin-up for the future.

If folks want the kNN implementations (if, for instance, they have a theoretical reason to want to use kNN with some odd distance metric, or they want to look into kNN on random points instead of grids for real problems to test the Bergstra & Bengio explanation for POMDPs), it is available in the archived pull request and in the source fork.

[coveralls]

Coverage decreased (-1.4%) to 91.266% when pulling e60cf4af46ec5c02f4db3f524d4389e4e7d699b4 on ptoman:add_tests into 3559c40362b5b12beaf9681f4a804b76592f794c on sisl:master.

[coveralls]

Coverage decreased (-1.4%) to 91.266% when pulling e60cf4af46ec5c02f4db3f524d4389e4e7d699b4 on ptoman:add_tests into 3559c40362b5b12beaf9681f4a804b76592f794c on sisl:master.

[zsunberg]

Thanks for contributing this @ptoman. The added tests are very helpful. I guess the coverage % decrease is because of adding the code to the benchmark functions in src, which is fine.

[ptoman]

I'm pretty sure the test coverage decreased because I commented out some of the benchmarking tests. I just uncommented them and re-pushed to the fork. Should I submit another pull request?

[zsunberg]

Oh yeah sure, that would be great!

On Thu, Jan 12, 2017, 15:16 Pamela Toman notifications@github.com wrote:

I'm pretty sure the test coverage decreased because I commented out some of the benchmarking tests. I just uncommented them and re-pushed to the fork. Should I submit another pull request?

— You are receiving this because you modified the open/close state.

Reply to this email directly, view it on GitHub https://github.com/sisl/GridInterpolations.jl/pull/10#issuecomment-272314633, or mute the thread https://github.com/notifications/unsubscribe-auth/AEC0axPA4D8m_Ymv8TKmcKHuVCIp05Duks5rRrRbgaJpZM4LiU4Y .