ML Evolution Agent 🤖

Auto-evolving ML competition agent that learns from every experiment.

What This Skill Does

1. Auto-evolves ML models for Kaggle-style competitions
2. HCC Multi-layer Memory - Episodic, Pattern, Knowledge, Strategic layers
3. Continuous improvement - Each phase learns from previous failures/successes
4. Resource-aware - Respects system limits (time, memory, API quotas)

When to Use

• User mentions Kaggle competition
• Tabular data classification/regression tasks
• Need to beat a target LB score
• User wants automated ML experimentation

Quick Start

# Initialize
from ml_evolution import MLEvolutionAgent

agent = MLEvolutionAgent(
    competition="playground-series-s6e2",
    target_lb=0.95400,
    data_dir="./data"
)

# Run evolution
agent.evolve(max_phases=10)

HCC Memory Architecture

Layer 1: Episodic Memory
├── Experiment logs (phase, CV, LB, features, params)
├── Success/failure records
└── Resource usage tracking

Layer 2: Pattern Memory
├── What works (success patterns)
├── What fails (failure patterns)
└── When to use each approach

Layer 3: Knowledge Memory
├── Feature engineering techniques
├── Model configurations
├── Hyperparameter knowledge
└── Domain-specific features

Layer 4: Strategic Memory
├── Auto-evolution rules
├── Resource management rules
├── Exploration-exploitation balance
└── Competition-specific strategies

Proven Techniques (from real competitions)

Feature Engineering

Model Configurations

Resource Limits

• Features: < 60 (avoids timeout)
• Iterations: < 1200 (avoids SIGKILL)
• Training time: < 20 min (system limit)
• Submissions: 10/day (Kaggle quota)

Evolution Rules

# Auto-evolution decision tree
if phase_improved:
    keep_features()
    try_similar_approach()
elif phase_degraded > 0.0001:
    rollback()
    try_new_direction()
else:
    fine_tune_params()

# Overfitting detection
if cv_lb_gap > 0.002:
    increase_regularization()
    reduce_features()
    simplify_model()

Files Structure

ml-evolution-agent/
├── SKILL.md              # This file
├── HCC_MEMORY.md         # Memory architecture details
├── FEATURE_ENGINEERING.md # Feature techniques library
├── MODEL_CONFIGS.md      # Optimal model configurations
├── EVOLUTION_RULES.md    # Auto-evolution decision rules
└── templates/
    ├── train_baseline.py # Baseline training script
    ├── train_evolved.py  # Evolution training script
    └── memory.json       # Example memory state

Example Results

Playground S6E2 (Feb 2026)

• Started: LB 0.95347
• Best: LB 0.95365 (+0.00018)
• Phases: 14
• Success rate: 36%
• Target beaten: Yes (0.95361 → 0.95365)

Key Learnings

1. Simple > Complex - Target stats beat complex feature engineering
2. Resource limits matter - Too many features = timeout
3. CatBoost is king - Consistently best for tabular data
4. Daily quota awareness - Kaggle limits submissions

Installation

clawhub install ml-evolution-agent

Built from real competition experience. Evolved through 14 phases of experimentation.