Skip to content

TuRBO (region slot)

Cite: Eriksson et al., NeurIPS 20191.

Paper Code

Switch with lenz set-region --policy turbo. Subsequent suggest calls (without --bounds / --around) optimize the acquisition function inside a hyperrectangle trust region. The region is centered on the incumbent. Length doubles after a streak of improvements and halves after a streak of failures. Below a minimum length the region restarts.

You own mode (enable, disable, override). TuRBO owns the counters.

lenz turbo init --state ./state.json
lenz turbo status --state ./state.json
lenz turbo override --state ./state.json --length 0.4 --center '{"x": 0.3}'
lenz set-region --state ./state.json --policy box

Prefer this over improvised set-bounds when you want principled local search. Still use suggest --around for a one-shot local probe that does not persist.

  1. Eriksson, Pearce, Gardner, Turner, Poloczek. Scalable Global Optimization via Local Bayesian Optimization. NeurIPS 2019. ↩