← Back to Publications

Genetic Algorithms with DNN-Based Trainable Crossover as an Example of Partial Specialization of General Search

Authors:

Year: 2017
Venue: arXiv:1809.04520
Links: arXiv abstract

Summary

Proposes partial specialization of general (universal-induction) search via genetic algorithms with a trainable crossover operator implemented as a deep feedforward neural network. A feasibility study shows that such specialized GAs can be more efficient than both general GAs and fully discriminative, search-free models, balancing generality and efficiency.

Relevance to Hyperon

An evolutionary-search source relevant to MOSES's estimation-of-distribution / evolutionary-programming core — it explores learnable crossover operators as a way to specialize general program search, adjacent to MOSES's meta-optimizing approach in the MOSES Deep Dive.

Key References

• RawData source
• MOSES Deep Dive