Best-of-N
Give the LLM N valid attempts at the same parent before committing to the global best, then repeat.
# Best-of-N — faithful port of SkyDiscover's BestOfNDatabase defaults.
# Sections mirror the six core components.
seed: 0
general:
max_iterations: 100
inner_retry_times: 3 # SkyDiscover retries up to 3 propose+evaluate attempts per iteration (retry_times)
# population (BestOfNPopulation) is uncapped keep-all by default (SkyDiscover enforces no cap);
# set population.capacity to an int only if you need to bound the archive.
selection_policy: # BestOfNPolicy
best_of_n: 5 # reuse the same parent for N consecutive attempts before switching to best
num_inspirations: 4 # context programs sampled from the top pool (num_context_programs)
proposer:
max_tokens: 32000 # SkyDiscover LLMConfig.max_tokens (applies only to a model built from this preset)