recipe_ai_hrm_v1

RecipeAI: Hierarchical Recipe Generation with RL

Model Description

This is a trained PPO agent for nutritionally-balanced recipe generation, part of a hierarchical reinforcement learning system (HRM).

Phase 1 Baseline Agent:

• Algorithm: Proximal Policy Optimization (PPO)
• Training: 150,000 timesteps
• Framework: Stable Baselines3
• Environment: Custom Gymnasium env (RecipeEnv)

Performance

Phase 1 (Standalone)

• Constraint Compliance: 100% (deterministic mode)
• Recipe Diversity: 1% (deterministic), 96% (stochastic)
• Average Reward: -164.06

Phase 2 (HRM Integration)

• Weekly Target Achievement: 4/5 nutrients within 10%
• Recipe Diversity: 100% unique recipes
• Unique Ingredients: 55/324 used
• Daily Compliance: 14% (expected due to dynamic constraint adjustment)

Training Details

• State Space: 11-dimensional

  • Current nutrients (5): calories, protein, sodium, carbs, fat
  • Target nutrients (5)
  • Ingredient count (1)

• Action Space: 325 discrete actions

  • 324 ingredients
  • 1 DONE action

• Constraints (per meal):

  • Calories: 400-800 kcal
  • Protein: 15-50g
  • Sodium: 0-800mg
  • Carbs: 30-100g
  • Fat: 10-30g

• Dataset: USDA FoodData Central (324 processed ingredients)

Usage

import gymnasium as gym
from stable_baselines3 import PPO

# Load model
model = PPO.load("recipe_agent_standard_ppo")

# Generate recipe (requires RecipeEnv - see repository)
# obs, info = env.reset()
# action, _states = model.predict(obs, deterministic=True)

Repository

Full code: [RecipeAI GitHub Repository]

Citation

@software{recipeai2024,
  title={RecipeAI: Hierarchical Reinforcement Learning for Recipe Generation},
  author=Bhavesh,
  year=2025,
  url={https://huggingface.co/bhxvxsh/recipe-ai-hrm}
}

License

MIT License