Work

Cross-representation mechanics of genetic-programming search, Stage 4: do restarts close genetic programming's gap to exact blind search in symbolic regression? (H11)

Year: 2026
Type: study-registration
Abstract: Stage 1 (OSF a3spg) confirmed that in real-valued symbolic regression one GP run reaches smaller targets than size-ordered enumeration of distinct functions (the exact blind-search null, whose reach is log2 E bits with E evaluations): 4.13 bits less at E0 = 20,000 evaluations. Stage 3 (OSF hb238) confirmed why: without guidance GP loses about the same to blind search in both domains studied, and in the real domain guidance is small (about 3 bits at E0) and does not grow with budget along a run. A fresh run re-earns guidance on fresh coverage, so independent restarts should turn budget into reach faster than one long run. Stage 1's allocation runs point the same way. This Stage 4 tests the prediction on the six real held-out function sets of Stage 2 (OSF dg7f3) with new runs: GP restarted every 100 generations, against exact blind search and against one long run, at 16 x E0.
Language: English
Open Access status: green
OpenAlex ID: W7225102518
Cites:
Cited by:
Related to: 10