AutoLOGIC README

The project's README from github.com/sethGu/AutoLOGIC. As one markdown file: README.md.

AutoLOGIC

AutoLOGIC is an agentic framework for automated tabular learning across classification, regression, and clustering. It converts a task specification into coordinated feature, model, optimization, and output-control plans, executes the plans with task-specific tools, and records plan revisions from validation and execution observations.

Agentic workflow

  1. TaskSpec defines the task, dataset, primary metric, output requirements, and execution constraints.
  2. AgenticReasoningAgent coordinates four plan channels: feature engineering, model construction, optimization, and output control.
  3. Task-specific executors generate and evaluate feature programs, candidate models, and parameter configurations.
  4. Validation scores, failures, retries, and fallbacks are returned to the agent as observations.
  5. The agent records a retain, revise, or fallback transition for the affected plan channel.
  6. The selected route produces stacking, distillation, calibration, prediction-interval, or clustering outputs according to the task.

Repository structure

Installation

python -m venv .venv
python -m pip install -r requirements.txt

Configure data paths and the LLM endpoint through environment variables:

OPENAI_API_KEY
OPENAI_BASE_URL
AUTOLOGIC_DATA_DIR
AUTOLOGIC_CSV_DATA_DIR
AUTOLOGIC_REGRESSION_DATA_DIR

Task runners

python autologic/ensemble/classification_ensemble/classification_auto_ensemble.py --help
python autologic/ensemble/multiclassification_ensemble/multiclassification_auto_ensemble.py --help
python autologic/ensemble/regression_ensemble/regression_auto_ensemble.py --help
python autologic/ensemble/cluster_ensemble/cluster_auto_ensemble_hidden_evaluator.py --help

Each runner accepts --task-spec <path.json>. The generated protocol file contains the task specification, four plans, agent configuration, and execution transitions.

python autologic/ensemble/classification_ensemble/classification_auto_ensemble.py --task-spec configs/task_spec.classification.example.json

Figure reproduction

pwsh reproduce/fig4/run_full_benchmark.ps1

Fig. 5 and Fig. 6 scripts are organized by panel under reproduce/fig5/ and reproduce/fig6/.

Source verification

python scripts/check_release.py
python scripts/build_manifest.py

AutoLOGIC (sethGu/AutoLOGIC)

Searching Database…

这是什么?AutoLOGIC 是一个面向自动化表格机器学习的智能体框架,覆盖分类、回归和聚类。它把任务规范转化为相互协调的特征、模型、优化和输出控制计划,然后用任务专用工具执行这些计划。
这是开箱即用的工具,还是要写脚本对接的研究框架?它是一个研究框架:你把任务描述为 TaskSpec 配置,并用 --task-spec 启动对应的 Python 执行器,不是点点鼠标就能用的应用。
怎么让一个任务跑起来?通过环境变量设置数据集路径和 LLM 端点,然后用 --task-spec 指向你的配置文件来运行任务专用执行器。
使用它需要 LLM API 吗?需要,AgenticReasoningAgent 依赖一个已配置的 LLM 端点,与数据路径一起通过环境变量设置。
它是生产可用的,还是更像研究产物?它看起来是为论文配套的研究代码:它带有单独的 reproduce/fig4、fig5、fig6 脚本用于重新生成图表,并依赖 legacysage/compat 基准模块,而不是为生产使用而打包。
agent 协调的四个计划通道是什么?它协调特征工程、模型构建、优化和输出控制计划,每一项都根据执行反馈独立修订。
一个任务能产出哪些类型的输出?根据任务不同,AutoLOGIC 可以产出堆叠、蒸馏、校准、预测区间或聚类输出。
agent 如何决定是否更改计划?执行之后,验证得分、失败、重试和回退会作为观察结果返回,agent 会为受影响的计划通道记录一次保留、修订或回退的转换。
我能用这个仓库复现论文的图表吗?可以,reproduce/fig4、fig5 和 fig6 文件夹包含用于生成那些图表的基准执行、校准、蒸馏和聚类脚本。
Creator
sethGu
Size
0.4 MB
Stars
209
Updated
CommitSep 12
Age
1mo
Commits · 90 Days