Deviations from Duvenaud

This page lists and explains some of the deviations this project does (on purpose) from the one it tries to improve.

Deviations

Re-adding best Kernel from previous Iteration

Kernels from previous iteration (#19) is not done on purpose, as:

• AS never seems to do this (best_models is never initialized),
• the search operators described in Duvenauds Thesis, Appendix C and implemented in _grammar_duvenaud.py never allow this.

Also it doesn't seem to make any sense with the current optimizer (ScipyOptimizer), as minimization should yield deterministic results (up to a small error, random restarts).