Wanted algorithms

[ffelten]

This is a list of algorithms we would like to include and use in EngiOpt.

• 3D CGAN for HeatConduction (in progress) closed in 3D GANs  #43
• VQGAN (in progress @arthurdrake1 ) closed in Vqgan #49
• Least volume (in progress @mkeeler43)
• Autoregressive model (in progress) see Final version of the PixelCNN++ implementation #51
• Image conditioned model
• Super Simple Baseline: Just pick the closest design in the training dataset to the requested condition (i.e., KNN w/ K=1) -- Came up in a reviewer rebuttal and was surprisingly effective, so could be a useful simple baseline.

[arthurdrake1]

@ffelten Just to notify that I am starting on VQGAN today. Will update as I develop the model.

[mkeeler43]

Least Volume is completed as well. I am just not sure we can go public with the generalized variant yet if Qiuyi still plans on publishing. Maybe @markfuge can comment on this

[markfuge]

@mkeeler43 We are waiting on ArXiV review and approval presently as he submitted it a month ago but there was a manual review step needed. I can update when this is cleared to go

[markfuge]

FYI @mkeeler43 the paper just went live on ArXiV today: https://arxiv.org/abs/2404.17773 It is under review presently at a journal, but since the ArXiV is out now, I think it is OK to move forward.

[markfuge]

FYI @ffelten I added another baseline to the wanted algorithms list based on a useful comment a reviewer made, which is to just compare against picking the closest sample in the training dataset to the requested condition, and just using that sample. This is effectively KNN with K=1, which Milad also found worked well in the JMD Mean Squared Error paper. Did well on comparative tests on Beams2D and would be a fun and easy one to include as a baseline method since it is super simple and more performant than random.

[arthurdrake1]

After lots of testing and code review, VQGAN is now complete and merged! Thanks to @mkeeler43 and @ffelten for their help 🚀