$E^3$-ML-Master: Advanced Envisioning-Executing for Goal-Driven Self-Evolved ML-Master

Project Overview

Note

Derived from ML Agent: ML-Master

E³-ML-Master is a long-horizon, self-evolved ML agent framework that uses Monte Carlo Tree Search (MCTS) and LLM-driven code generation to solve machine learning competition tasks. The framework employs a dual-core design: Envisioner (global exploration decision maker) and Executor (parallel execution engine).

Core Components

• Envisioner (framework/agent.py): Maintains the global MCTS search tree, coordinates multiple parallel Executors, manages global Memory system, tracks global best node
• Executor (framework/agent.py): Executes strategy code for specific nodes, uses LLM to extract metrics from execution output, supports multi-turn refinement for code optimization
• MCTSNode (framework/node.py): Represents a strategy in the search tree with fields for visits, expansion, rewards, and UCT calculation
• Memory (framework/memory.py): Context-aware memory system storing all exploration history with indexed retrieval
• Interpreter (interpreter/interpreter_parallel.py): Parallel code execution engine that runs Python code in isolated environments
• Backend (backend/): LLM API interface layer supporting OpenAI-compatible and Qwen backends

Quick StartUp

Environment SetUp

uv sync
source .venv/bin/activate
# Python 3.13.5

APIKEY SetUp

By default, we use deepseek-v3.2 (deepseek-chat) for the backbone of Executor Agent for writing codes and instruction followings, and we use deepseek-r1 (deepseek-reasoner) for the backbone of core Envisioner Agent for self-reflection and expanding new strategies.

• OPENAI_API_KEY with deepseek backbone: Official API
• SERPER_API_KEY for web search and web parse tool calling. SERPER_DEV

Write the following api keys into .env file.

OPENAI_API_KEY="Your key"
BASE_URL="https://api.deepseek.com"
SERPER_API_KEY="Your key"

Usage

python main.py

Detailed Workflow

Important

All the source code of the new architecture is in framework folders.