Comparison of Generative Meta-Learning vs Nevergrad
on the 30-dimensional Schwefel function optimization
The best results of both methods after 100K trials:
gen-meta best_epoch: 99500 loss: 1.597656 time: 1.372849
ng-opt-4 best_epoch: 67590 loss: 476.789062 time: 63.584929
average gen-meta loss after 10 repetitions: 233.1332031
average ng-opt-4 loss after 10 repetitions: 409.2259765
Please note that, an experiment with several random
seeds is required to correctly compare both of them.
Solving math functions in high-dimensions: gen_meta_100k.py
Matrix Factorization on MovieLens 1M dataset: gen_matrix.py
f1@10: 86% ncdg@10: 60% f1@100: 79% ncdg@100: 40%
Selecting portfolios for sparse index tracting (vs Fast CMA-ES):
https://github.com/kayuksel/generative-opt