人工智能大模型AI Agent代碼智能體自主智能體【免費下載鏈接】openevolveOpen-source implementation of AlphaEvolve項目地址https://gitcode.com/gh_mirrors/op/openevolve點擊查看免費下載OpenEvolve 是一個開源的 AlphaEvolve 實現(xiàn)它把 LLM 變成自主代碼優(yōu)化器通過進化的方式反復(fù)改進程序。本篇文章以倉庫configs/目錄下的配置體系為核心系統(tǒng)講解如何從default_config.yaml全量模板出發(fā)理解每一個配置參數(shù)的含義與默認值掌握島嶼模型Island-Based Evolution三個核心參數(shù)num_islands、migration_interval、migration_rate的調(diào)優(yōu)邏輯并基于island_examples.yaml的五大典型場景與early_stopping_example.yaml的早期停止示例寫出可直接運行的實戰(zhàn)配置。讀完本文你將具備從零搭建 OpenEvolve 進化任務(wù)配置、按問題復(fù)雜度選擇合適的島嶼策略、并通過命令行或 Python API 一鍵啟動的能力。configs 目錄OpenEvolve 配置文件的完整圖譜configs/目錄是 OpenEvolve 的配置中心共包含 4 個 YAML 文件與一份導(dǎo)讀文檔configs/README.md它們分工明確文件定位configs/default_config.yaml主配置模板包含所有可用選項與合理的默認值參數(shù)注釋最全適合作為自定義配置的起點configs/island_config_example.yaml島嶼模型實戰(zhàn)示例演示推薦的海島設(shè)置、平衡的遷移參數(shù)是一份可直接運行的完整配置configs/island_examples.yaml五種不同場景的島嶼配置集合最大多樣性、聚焦探索、均衡、快速探索、大規(guī)模進化并附參數(shù)選擇準則configs/early_stopping_example.yaml早期停止功能示例演示如何在進化的同時自動判斷是否已無改進空間其中default_config.yaml是靈魂文件它既是一份參數(shù)字典也是 OpenEvolve 各核心模塊LLM、Prompt、Database、Evaluator、EvolutionTrace配置入口的真實寫照與 openevolve/config.py 中的Config及其子 dataclassLLMConfig、PromptConfig、DatabaseConfig、EvaluatorConfig、EvolutionTraceConfig一一對應(yīng)。下面我們先逐模塊拆解這份全量模板。default_config.yaml全量配置模板逐模塊拆解通用設(shè)置General Settingsmax_iterations: 100 # Maximum number of evolution iterations checkpoint_interval: 10 # Save checkpoints every N iterations log_level: INFO # Logging level (DEBUG, INFO, WARNING, ERROR, CRITICAL) log_dir: null # Custom directory for logs (default: output_dir/logs) random_seed: 42 # Random seed for reproducibility (null random, 42 default)max_iterations進化總迭代數(shù)上限。注意這是 YAML 模板中的示例值代碼層Config的默認值為 10000見 openevolve/config.pyCLI 也可通過--iterations覆蓋checkpoint_interval每隔 N 次迭代保存一次檢查點配合 CLI 的--checkpoint參數(shù)可實現(xiàn)斷點續(xù)跑random_seed設(shè)置后控制器會同步 seed 全局random、numpy并對不同組件派生哈希種子保證可復(fù)現(xiàn)設(shè)為null則隨機。進化設(shè)置Evolution Settingsdiff_based_evolution: true # Use diff-based evolution (true) or full rewrites (false) max_code_length: 10000 # Maximum allowed code length in characters enforce_evolve_blocks: false # Revert LLM edits outside EVOLVE-BLOCK-START/END markersdiff_based_evolutiontrue采用基于 SEARCH/REPLACE diff 的增量進化false則使用全量重寫。Config中還維護diff_pattern正則默認 SEARCH\n(.*?)\n(.*?) REPLACE用于解析 LLM 返回的 diffenforce_evolve_blocks開啟后LLM 在EVOLVE-BLOCK-START/EVOLVE-BLOCK-END標記之外的編輯會被回滾實現(xiàn)位于 openevolve/utils/code_utils.py 的enforce_evolve_blocks。早期停止設(shè)置Early Stoppingearly_stopping_patience: null # Stop after N iterations without improvement (null disabled) convergence_threshold: 0.001 # Minimum improvement required to reset patience counter early_stopping_metric: combined_score # Metric to track for early stopping詳見本文早期停止Early Stopping配置一節(jié)。LLM 配置llmllm: models: # List of available models with their weights - name: gemini-2.0-flash-lite weight: 0.8 - name: gemini-2.0-flash weight: 0.2 evaluator_models: # Models for LLM feedback - name: gemini-2.0-flash-lite weight: 0.8 - name: gemini-2.0-flash weight: 0.2 api_base: https://generativelanguage.googleapis.com/v1beta/openai/ # Base URL for API api_key: null # API key (defaults to OPENAI_API_KEY env variable) # api_key: ${GEMINI_API_KEY} # Reads API key from $GEMINI_API_KEY temperature: 0.7 # Higher more creative top_p: 0.95 # Top-p sampling parameter max_tokens: 4096 # Maximum tokens to generate timeout: 60 # Timeout for API requests in seconds retries: 3 # Number of retries for failed requests retry_delay: 5 # Delay between retries in seconds幾個值得注意的源碼級細節(jié)多模型加權(quán)集成models列表支持配置多個模型并賦予權(quán)重運行時由LLMEnsemble按權(quán)重采樣見 openevolve/llm/ensemble.py。evaluator_models若未顯式配置會復(fù)用models列表openevolve/config.py${VAR}環(huán)境變量語法api_key支持api_key: ${GEMINI_API_KEY}形式配置加載時會通過_resolve_env_var完整匹配并展開若引用的環(huán)境變量未設(shè)置會拋出ValueErroropenevolve/config.py向后兼容語法除了models數(shù)組還兼容primary_model/primary_model_weight與secondary_model/secondary_model_weight的舊寫法LLMConfig.__post_init__會自動把它們轉(zhuǎn)換為models列表條目openevolve/config.pyisland_config_example.yaml與 examples/circle_packing/config_phase_1.yaml 均使用了這種寫法provider 擴展LLMModelConfig還支持provider字段openai默認、claude_code、copilot_cli以及reasoning_effort、max_budget_usdClaude Code、max_ai_creditsCopilot CLI、manual_mode人機協(xié)同等進階選項。Prompt 配置promptprompt: template_dir: null # Custom directory for prompt templates system_message: You are an expert coder helping to improve programs through evolution. island_system_messages: [] # Optional literal text or template names by island index evaluator_system_message: You are an expert code reviewer. num_top_programs: 3 # Number of top-performing programs to include num_diverse_programs: 2 # Number of diverse programs to include use_template_stochasticity: true # Use random variations in templates for diversity template_variations: # Different phrasings for parts of the template improvement_suggestion: - Heres how we could improve this code: - I suggest the following improvements: - We can enhance this code by: include_artifacts: true # Include execution outputs/errors in prompt max_artifact_bytes: 20480 # Maximum artifact size in bytes (20KB default) artifact_security_filter: true # Apply security filtering to artifacts suggest_simplification_after_chars: 500 include_changes_under_chars: 100 concise_implementation_max_lines: 10 comprehensive_implementation_min_lines: 50 diff_summary_max_line_len: 100 diff_summary_max_lines: 30template_dir自定義提示詞模板目錄。加載時若為相對路徑會相對配置文件所在目錄解析openevolve/config.py倉庫默認模板見 openevolve/prompts/defaults/island_system_messages按島嶼下標為不同島嶼指定專屬系統(tǒng)消息字面量或模板名null/缺失則回退到system_message。它必須與database.num_islands配合條目數(shù)不能超過島嶼數(shù)否則Config.from_dict會拋出ValueErroropenevolve/config.py閾值類參數(shù)suggest_simplification_after_chars等用于特征提取與程序標簽簡潔實現(xiàn)/全面實現(xiàn)diff_summary_*控制 Previous Attempts 段落中 SEARCH/REPLACE 塊的展示長度。Database 配置databasedatabase: db_path: null # Path to persist database (null in-memory only) in_memory: true log_prompts: true population_size: 1000 # Maximum number of programs to keep in memory archive_size: 100 # Size of elite archive num_islands: 5 # Number of islands (separate populations) migration_interval: 50 # Migrate between islands every N generations migration_rate: 0.1 # Fraction of top programs to migrate (0.1 10%) elite_selection_ratio: 0.1 # Ratio of elite programs to select exploration_ratio: 0.2 # Ratio of exploration vs exploitation exploitation_ratio: 0.7 # Ratio of exploitation vs random selection feature_dimensions: # MAP-Elites feature map dimensions - complexity # Code length (built-in) - diversity # Code diversity (built-in) feature_bins: 10 # Number of bins per dimension diversity_reference_size: 20 # Size of reference set for diversity calculationdb_path持久化路徑null時數(shù)據(jù)庫僅在內(nèi)存中運行MAP-Elites 特征圖feature_dimensions默認[complexity, diversity]兩者均為內(nèi)置特征代碼長度與結(jié)構(gòu)多樣性也可混入評估器返回的自定義指標如performance、correctness。注意自定義維度必須由評估器返回連續(xù)的原始數(shù)值OpenEvolve 內(nèi)部完成縮放與分箱評估器不應(yīng)返回預(yù)先算好的 bin 索引feature_bins每個維度的分箱數(shù)可以是整數(shù)所有維度一致或按維度指定的字典例如feature_bins: {complexity: 10, diversity: 15}從源碼看diversity_metric目前固定為edit_distancefeature_based 未實現(xiàn)選擇比例默認elite_selection_ratio0.1、exploration_ratio0.2、exploitation_ratio0.7。Evaluator 配置evaluatorevaluator: timeout: 300 # Maximum evaluation time in seconds max_retries: 3 # Maximum number of retries for evaluation cascade_evaluation: true # Use cascade evaluation to filter bad solutions early cascade_thresholds: # Thresholds for advancing to next evaluation stage - 0.5 - 0.75 - 0.9 parallel_evaluations: 4 # Number of parallel evaluations use_llm_feedback: false # Use LLM to evaluate code quality (experimental) llm_feedback_weight: 0.1 # Weight for LLM feedback in final score適應(yīng)度計算優(yōu)先使用評估器返回的combined_score否則對所有指標取平均但會排除database.feature_dimensions中列出的特征維度級聯(lián)評估cascade先跑廉價的第一階段達到閾值0.5才進入下一階段用于盡早過濾劣質(zhì)解對應(yīng) openevolve/evaluator.py 中的_cascade_evaluate與run_stage1/2/3資源限制memory_limit_mb、cpu_limit與分布式評估distributed在注釋中明確標注not yet implemented配置時不要依賴。Evolution Trace 配置evolution_traceevolution_trace: enabled: false # Enable evolution trace logging format: jsonl # Output format: jsonl, json, or hdf5 include_code: false # Include full program code in traces include_prompts: true # Include prompts and LLM responses output_path: null # Defaults to output_dir/evolution_trace.{format} buffer_size: 10 # Number of traces to buffer before writing compress: false # Compress output file (jsonl only)該模塊為強化學(xué)習(xí)訓(xùn)練與進化分析記錄細粒度的進化軌跡支持 jsonl/json/hdf5 三種格式與可選壓縮底層實現(xiàn)在 openevolve/evolution_trace.py。島嶼模型OpenEvolve 保持進化多樣性的核心機制configs/README.md明確指出default_config.yaml的價值不僅在于參數(shù)齊全還在于內(nèi)置了島嶼式進化參數(shù)Island-Based Evolution Parameters——這是 OpenEvolve 維持進化多樣性的關(guān)鍵設(shè)計。島嶼模型的思想是不維護一個全局種群而是把種群拆成多個獨立進化的子種群島嶼每個島嶼在自己的搜索區(qū)域深耕每隔若干代把各島嶼上的優(yōu)秀程序遷移到相鄰島嶼實現(xiàn)知識共享。這樣既避免了單一種群過早收斂到局部最優(yōu)又通過周期性交流保留了全局進步的動力。三個核心參數(shù)database: num_islands: 5 # Number of separate populations migration_interval: 50 # Migrate every N generations migration_rate: 0.1 # Fraction of top programs to migratenum_islands島嶼數(shù)量。島嶼越多獨立搜索方向越多、多樣性越強但收斂變慢推薦大多數(shù)問題取 3–10migration_interval每隔多少代遷移一次。值越大各島嶼越獨立推薦 25–100 代migration_rate每次遷移時從每個島嶼頂部程序按適應(yīng)度降序中遷出的比例推薦 0.05–0.25%–20%。比例越高知識傳播越快。源碼級的遷移實現(xiàn)細節(jié)在 openevolve/database.py 中可以看到島嶼機制的真實實現(xiàn)輪轉(zhuǎn)調(diào)度next_island()以 round-robin 方式切換當(dāng)前進化島嶼increment_island_generation()維護每個島嶼的代數(shù)計數(shù)器openevolve/database.py遷移判定should_migrate()比較最大島嶼代數(shù)與上次遷移代數(shù)的差值是否達到migration_intervalopenevolve/database.py遷移數(shù)量migrate_programs()將每個島嶼的程序按get_fitness_score優(yōu)先combined_score否則對非特征維度指標取平均降序排列取max(1, int(len(island_programs) * migration_rate))個作為遷出對象openevolve/database.py環(huán)狀拓撲默認把遷出程序復(fù)制到相鄰島嶼(i1) % n與(i-1) % nopenevolve/database.py即鄰居共享而非全局廣播去重保護_migrate_one()會跳過目標島嶼已存在相同代碼的程序并為遷移副本生成新 UUID、打上migrant: true標記openevolve/database.py最少島嶼限制島嶼數(shù)小于 2 時直接跳過遷移openevolve/database.py。這些細節(jié)解釋了為什么migration_rate是比例而非數(shù)量遷移量隨島嶼種群規(guī)模動態(tài)縮放小種群也能保證至少遷移 1 個程序。何時使用什么參數(shù)選擇速查configs/README.md與 configs/island_examples.yaml 給出的經(jīng)驗法則問題特征策略說明復(fù)雜問題易陷入局部最優(yōu)更多島嶼 更低遷移頻率增大搜索空間覆蓋延長各島獨立演化時間簡單問題結(jié)構(gòu)清晰更少島嶼 更頻繁遷移加速知識共享快速收斂長時運行更多島嶼用島嶼多樣性對抗長跑中的早熟收斂短時運行更少島嶼更快收斂避免資源浪費island_config_example.yaml一份可直接運行的島嶼配置configs/island_config_example.yaml 是島嶼模型的標準答案它同時示范了兩套 LLM 配置語法與島嶼專屬系統(tǒng)消息的用法max_iterations: 1000 checkpoint_interval: 100 log_level: INFO llm: primary_model: gemini-2.0-flash-lite primary_model_weight: 0.8 secondary_model: gemini-2.0-flash secondary_model_weight: 0.2 temperature: 0.7 top_p: 0.95 max_tokens: 4096 database: population_size: 500 archive_size: 100 num_islands: 5 # 5 個獨立種群 migration_interval: 50 # 每 50 代遷移 migration_rate: 0.1 # 遷移頂部 10% 的程序 elite_selection_ratio: 0.1 exploration_ratio: 0.3 exploitation_ratio: 0.7 feature_bins: 10 prompt: system_message: Improve the program through evolution. island_system_messages: - Focus on simpler algorithms. - null - Focus on runtime performance. num_top_programs: 3 num_diverse_programs: 2 use_template_stochasticity: true evaluator: timeout: 300 max_retries: 3 cascade_evaluation: true parallel_evaluations: 4 diff_based_evolution: true allow_full_rewrites: false max_code_length: 10000亮點解讀島嶼專屬提示詞island_system_messages允許給不同島嶼定制演化風(fēng)格——島嶼 0 專注更簡單的算法、島嶼 2 專注運行時性能第二個條目為null表示該島嶼沿用system_message。這種分工進化正是島嶼模型多樣性的延伸兼容語法primary_modelsecondary_model的寫法會在Config加載時被自動展開為models權(quán)重列表主模型 0.8、次模型 0.2feature_bins 支持字典形式注釋中展示了feature_bins: {performance: 20, correctness: 10}的按維度分箱寫法。island_examples.yaml五大典型場景配置速查configs/island_examples.yaml 把島嶼參數(shù)壓縮成五個可直接套用的場景模板# 1. 最大多樣性適合復(fù)雜問題、長時運行、避免局部最優(yōu) max_diversity: database: num_islands: 10 migration_interval: 25 migration_rate: 0.2 population_size: 1000 archive_size: 200 # 2. 聚焦探索適合結(jié)構(gòu)清晰的問題、較短運行 focused_exploration: database: num_islands: 3 migration_interval: 100 migration_rate: 0.05 population_size: 500 archive_size: 50 # 3. 均衡方案大多數(shù)問題的默認起點 balanced: database: num_islands: 5 migration_interval: 50 migration_rate: 0.1 population_size: 1000 archive_size: 100 # 4. 快速探索小規(guī)模、原型驗證 quick_exploration: database: num_islands: 3 migration_interval: 20 migration_rate: 0.15 population_size: 200 archive_size: 30 # 5. 大規(guī)模進化復(fù)雜優(yōu)化、超長進化運行 large_scale: database: num_islands: 15 migration_interval: 75 migration_rate: 0.08 population_size: 2000 archive_size: 300從 5 個場景可以觀察出清晰的權(quán)衡規(guī)律max_diversity用 10 個島嶼 高頻高比例遷移25 代 / 20%最大化搜索空間覆蓋focused_exploration用 3 個島嶼 低頻低比例遷移100 代 / 5%讓每個島嶼深入挖掘large_scale則把島嶼數(shù)推到 15、種群擴到 2000同時用保守的 8% 遷移率避免大規(guī)模種群被過早同化。參數(shù)選取的本質(zhì)是在多樣性多島、少遷與收斂速度少島、多遷之間找平衡點。早期停止Early Stopping配置configs/early_stopping_example.yaml 演示了如何在進化長跑中自動止損max_iterations: 1000 checkpoint_interval: 50 log_level: INFO # 連續(xù) 30 代無改進即停止 early_stopping_patience: 30 # Stop after 30 iterations without improvement convergence_threshold: 0.01 # Minimum improvement of 0.01 required to reset patience early_stopping_metric: combined_score # Track the combined_score metric llm: models: - name: gpt-4o-mini weight: 1.0 api_base: https://api.openai.com/v1 temperature: 0.7 max_tokens: 4096 database: population_size: 50 num_islands: 3 migration_interval: 20 evaluator: timeout: 60 max_retries: 2 parallel_evaluations: 2 diff_based_evolution: true max_code_length: 8000三個早期停止參數(shù)的語義見 openevolve/config.pyearly_stopping_patience連續(xù) N 次迭代無改進即停止null表示禁用。該示例中30表示30 代無改進就收工可顯著節(jié)省 LLM API 調(diào)用成本convergence_threshold判定有改進的最小幅度。只有當(dāng)early_stopping_metric的提升超過該閾值示例為 0.01才會重置耐心計數(shù)器若只是小幅抖動則不重置early_stopping_metric跟蹤的指標默認combined_score可改為評估器返回的其他指標名。該文件同樣是一個完整的可運行配置gpt-4o-mini OpenAI API適合預(yù)算有限、想快速驗證想法的場景。關(guān)于早期停止參數(shù)更細的行為與校驗邏輯可參考 tests/test_early_stopping_config.py。使用方法把配置跑起來方式一命令行啟動推薦快速上手先復(fù)制模板為自有配置cp configs/default_config.yaml my_config.yaml # Edit my_config.yaml for your specific needs然后通過倉庫入口腳本openevolve-run.py啟動python openevolve-run.py initial_program.py evaluator.py --config my_config.yaml --output my_outputCLI 支持的主要參數(shù)見 openevolve/cli.py參數(shù)說明initial_program位置參數(shù)初始程序文件路徑evaluation_file位置參數(shù)含evaluate函數(shù)的評估文件路徑--config/-c配置文件路徑Y(jié)AML--output/-o結(jié)果輸出目錄--iterations/-i覆蓋配置中的最大迭代數(shù)--target-score/-t達到目標分數(shù)即停止--log-level/-l覆蓋日志級別--checkpoint指定檢查點目錄續(xù)跑--api-base、--primary-model、--secondary-model覆蓋 LLM API 與模型配置實際項目中這種用法非常常見例如 examples/arc_benchmark/run_evolution.sh 中的啟動命令python ../../openevolve-run.py initial_program.py evaluator.py --config config.yaml --output $OUTPUT_DIR $方式二Python API 啟動適合集成與腳本化configs/README.md給出的 API 用法如下from openevolve import OpenEvolve from openevolve.config import Config config Config() config.from_yaml(my_config.yaml) evolve OpenEvolve( initial_program_pathprogram.py, evaluation_fileevaluator.py, configconfig )此外倉庫還提供了更簡潔的函數(shù)式 APIrun_evolutionopenevolve/api.py它接受文件路徑 / 代碼字符串 / 代碼行列表三種形式的初始程序以及文件路徑 / Callable兩種形式的評估器config參數(shù)既可以是 YAML 路徑也可以是Config對象from openevolve import run_evolution result run_evolution( initial_programprogram.py, evaluatorevaluator.py, configmy_config.yaml, iterations200, ) print(result.best_score)配置加載的底層行為了解即可從 openevolve/config.py 的Config.from_yaml/from_dict實現(xiàn)可知YAML 通過yaml.safe_load讀取再用dacite.from_dict轉(zhuǎn)換為類型化的Configdataclass 結(jié)構(gòu)因此字段名必須與 dataclass 定義一致llm.temperature/top_p為null時會被自動剔除避免 dacite 類型錯誤頂層random_seed會自動下放到database.random_seed兩處校驗會直接拋出ValueErrorprompt.island_system_messages條目數(shù)超過database.num_islandsprompt.programs_as_changes_descriptiontrue但diff_based_evolutionfalsediff_pattern若提供會先嘗試編譯正則非法表達式會報錯。從示例到實戰(zhàn)參考倉庫里的真實配置如果想看島嶼參數(shù)之外的完整實戰(zhàn)配置倉庫的 examples 目錄提供了大量真實任務(wù)的配置文件是絕佳的參考樣本examples/circle_packing/config_phase_1.yaml圓填充問題n26使用 4 個島嶼、cascade_thresholds: [0.5, 0.75]、diff_based_evolution: false構(gòu)造型程序用全量重寫以及含領(lǐng)域知識的長 system messageexamples/arc_benchmark/config.yamlARC 基準任務(wù)的配置配套post_evolution_eval.py做進化后評估examples/algotune/AlgoTune 基準下的多個算法優(yōu)化任務(wù)FFT、LU 分解、特征向量等每個任務(wù)都包含config.yaml、evaluator.py與initial_program.py三件套。綜合來看OpenEvolve 的配置哲學(xué)可以總結(jié)為三句話用default_config.yaml保證參數(shù)覆蓋完整用島嶼三參數(shù)num_islands/migration_interval/migration_rate控制多樣性與收斂的平衡用早期停止參數(shù)控制成本與收益的平衡。從復(fù)制模板、調(diào)整島嶼參數(shù)、按需開啟早期停止到用 CLI 或 API 啟動任務(wù)這就是一條完整且可復(fù)現(xiàn)的 OpenEvolve 配置實戰(zhàn)路徑。贊分享人工智能大模型AI Agent代碼智能體自主智能體【免費下載鏈接】openevolveOpen-source implementation of AlphaEvolve項目地址https://gitcode.com/gh_mirrors/op/openevolve點擊查看免費下載相關(guān)推薦CAI模型設(shè)置完全指南如何配置默認模型與單個代理CAICybersecurity AI是一個開源的人工智能平臺專為網(wǎng)絡(luò)安全測試和安全評估而設(shè)計。作為一名安全研究人員或滲透測試人員了解如何正確配置CAI人工智能AI Agent網(wǎng)絡(luò)安全滲透測試工具調(diào)用AI 評測2024最新three-globe入門教程5分鐘上手3D地球可視化開發(fā)2024最新three globe入門教程5分鐘上手3D地球可視化開發(fā) three globe是一款基于ThreeJS的WebGL地球數(shù)據(jù)可視化庫能幫助開發(fā)數(shù)據(jù)可視化3D渲染openclaw-cn models 命令完全指南模型發(fā)現(xiàn)、掃描、默認模型與認證配置openclaw cn models 命令完全指南模型發(fā)現(xiàn)、掃描、默認模型與認證配置 openclaw cn models 是中文社區(qū)版 OpenClawo人工智能AI Agent即時通訊后端本地部署語音上一篇Azure Community-Policy 與內(nèi)置策略對比何時選擇自定義解決方案下一篇解放雙手KeymouseGo自動化操作從入門到精通指南創(chuàng)作聲明：本文部分內(nèi)容由AI輔助生成（AIGC），僅供參考