这是一篇由原始材料转换而来的阅读页，保留了源文件的主要结构，并补充了可追溯的来源说明与链接。

Pattern: Search Strategy

Purpose

Define how the system explores the space of possible improvements instead of making arbitrary edits.

Core question

Given limited budget, what should the next experiment be?

Common strategies

1. Local hill-climbing

Small changes around the current best.

Use when: - the search space is smooth enough - interpretability matters - experiments are expensive

2. Staged exploration

Move from safer changes to more radical ones.

Example order: 1. hyperparameters 2. simplifications 3. local architecture tweaks 4. combined ideas 5. more radical restructuring

3. Bandit-style budget allocation

Spend more runs on promising directions and fewer on weak ones.

4. Parallel branch exploration

Different agents or branches explore different themes.

Helpful rules

• prefer one main hypothesis per run
• avoid mixing many unrelated changes at once
• revisit near-misses when new evidence appears
• exploit successful themes before wandering too far
• stop digging in a direction after repeated failures

Exploration heuristics

Useful heuristics include: - low-cost/high-signal changes first - simplifications early - combine only previously promising ideas - periodically re-read history before proposing next run

Anti-patterns

• random edits without hypothesis
• changing too many variables at once
• endlessly retrying a weak search direction
• ignoring history and repeating prior failures

Recommendation

Search strategy should be explicit enough that another agent could continue the loop coherently from the logs.