search-strategy.md

Search Strategy

openevolve/search_strategy.py — decides which candidate becomes the baseline for the next generation. Selected via create_strategy(config).

Strategies

flowchart TB
    H[candidate history] --> S{strategy}
    S -->|auto| A{stall = generations<br/>since last improvement}
    A -->|stall &lt; restart_patience| AG[greedy: best so far]
    A -->|restart_patience ≤ stall &lt; diversify_patience| AR[restart: gen-0 baseline]
    A -->|stall ≥ diversify_patience| AD[diversify: rotate top-K]
    S -->|greedy| G[pick the single best candidate]
    S -->|beam| B[pick randomly among top-K<br/>beam_width]
    S -->|random_restart| R{generation % restart_interval == 0?}
    R -->|yes| RB[revert to baseline<br/>escape local optima]
    R -->|no| RG[pick the best so far]
    A --> N([next baseline])
    G --> N
    B --> N
    RB --> N
    RG --> N

Loading

Comparison

Strategy	Selection	Parallelizable	Use when
auto (AutoEscalationSearch) default	Greedy, then escalates to restart → diversify based on plateau length	No	Zero-config; cheap by default, explores only when stuck
greedy (GreedySearch)	Always the top candidate	No	Fast convergence on smooth landscapes
beam (BeamSearch)	Random among top-beam_width	Yes	More exploration; parallel evaluation
random_restart (RandomRestartSearch)	Periodically revert to baseline every restart_interval	No	Escaping local optima

auto is deterministic — escalation is driven purely by stall (the number of generations since the last strict improvement in best score), so a rerun escalates at the exact same generation. It never spends extra LLM calls to decide, and because escalation only steers exploration, the loop still reports the highest-scoring candidate — auto never regresses below plain greedy.

Interface

classDiagram
    class SearchStrategy {
      <<abstract>>
      +select_baseline(history, generation) dict
      +should_parallelize() bool
    }
    class GreedySearch
    class BeamSearch {
      +int beam_width
    }
    class RandomRestartSearch {
      +int restart_interval
    }
    class AutoEscalationSearch {
      +int restart_patience
      +int diversify_patience
      +int beam_width
    }
    SearchStrategy <|-- GreedySearch
    SearchStrategy <|-- BeamSearch
    SearchStrategy <|-- RandomRestartSearch
    SearchStrategy <|-- AutoEscalationSearch

Loading

create_strategy({"strategy": "beam", "beam_width": 5}) returns the matching implementation; the OptimizerLoop calls select_baseline(history, generation) at the end of every generation. loopbench run defaults to auto; override it with --strategy or a search: block in loopbench.yaml (see Defining Your Benchmark).