CS4100-Group-Project

Meal Planning System (Simulated Annealing + Genetic Algorithm)

=================================================================

This project implements a daily-updating meal planning system that dynamically generates meal recommendations based on what the user has already eaten.

The system uses a rolling-horizon optimization framework, where:

• Meals already consumed are treated as fixed
• Remaining meals are optimized to meet nutritional goals

Two optimization methods are supported:

• Simulated Annealing (SA) – local search with probabilistic exploration
• Genetic Algorithm (GA) – population-based global search

A Gaussian Naive Bayes classifier (implemented from scratch with NumPy) is trained on the food dataset to predict user preference scores based on nutritional features, replacing hardcoded category-keyword rules.

---

Key Features

---

• Dynamic re-planning based on logged meals
• Multi-objective optimization including:
  • Nutrition targets (calories, protein, etc.)
  • Variety across meals
  • Meal compatibility (realistic combinations)
  • User preferences via Gaussian Naive Bayes (trained on nutritional features)
  • Follow-through modeling (foods the user tends to accept/reject)
• Greedy initialization for fast convergence
• Support for both SA and GA optimizers

---

Fitness Function

---

Each meal plan is evaluated using a weighted score: FinalFitness = w_nutrition * NutritionScore • w_variety * VarietyScore • w_compatibility * MealCompatibilityScore • w_preference * UserPreferenceScore • w_follow * FollowThroughScore

Where:

• NutritionScore: how close the plan is to macro targets
• VarietyScore: penalizes repeated categories
• MealCompatibilityScore: encourages realistic meal compositions
• UserPreferenceScore: predicted by a Gaussian Naive Bayes classifier trained on nutritional features; generalizes to foods in unknown categories
• FollowThroughScore: models whether the user is likely to follow the plan

---

Project Pipeline

---

1. Preprocess data

   • Combines USDA + OpenFoodFacts datasets
   • Cleans and limits dataset size

2. Greedy Initialization

   • Builds a strong starting meal plan

3. Optimization

   • SA refines a single solution
   • GA evolves a population of solutions

4. Final Output

   • Meal plan per slot
   • Full-day nutritional totals
   • Fitness breakdown

---

Setup Instructions

---

1. Install dependencies

uv sync

2. Download datasets

Download: • USDA FoodData Central (Foundation Foods) https://fdc.nal.usda.gov/fdc-datasets/FoodData_Central_csv_2025-12-18.zip • OpenFoodFacts dataset https://www.kaggle.com/datasets/openfoodfacts/world-food-facts Place them inside:

data/

such that the relative path of the FoodData dataset folder is data\FoodData_Central_foundation_food_csv_2025-12-18 and the OpenFoodFacts .tsv file is data\en.openfoodfacts.org.products.tsv

3. Preprocess data

uv run python src/preprocess.py

This generates:

data/processed/foods.csv

How to Run

---

Run with Simulated Annealing

uv run python main.py \
  --optimizer sa \
  --foods_path data/processed/foods.csv \
  --greedy_init \
  --slots breakfast lunch dinner snack \
  --slots_done breakfast

Run with Genetic Algorithm

uv run python main.py \
  --optimizer ga \
  --foods_path data/processed/foods.csv \
  --greedy_init \
  --slots breakfast lunch dinner snack \
  --slots_done breakfast

For Windows Computers:

uv run python main.py --optimizer sa --foods_path data/processed/foods.csv --greedy_init --slots breakfast lunch dinner snack uv run python main.py --optimizer sa --foods_path data/processed/foods.csv --greedy_init --slots breakfast lunch dinner snack --slots_done breakfast uv run python main.py --optimizer ga --foods_path data/processed/foods.csv --greedy_init --slots breakfast lunch dinner snack --slots_done breakfast

Important Arguments

•	--optimizer: sa or ga
•	--greedy_init: auto-generate starting plan (recommended)
•	--slots: all meal slots
•	--slots_done: meals already eaten
•	--foods_path: path to processed dataset

Optional: • nutrition targets (--target_energy, etc.) • scoring weights (--w_nutrition, etc.)

Example output

---

The program prints: • Meal plan per slot • Nutritional totals vs targets • Fitness breakdown

Example:

Full day totals (logged + planned)
  energy_kcal: 1907 / 2000
  protein_g: 50.1 / 50.0

Fitness Breakdown
  NutritionScore: 0.973
  VarietyScore: 1.000
  Final Fitness: 0.904

Notes & Limitations

---

•	Some meal combinations may be nutritionally optimal but unrealistic
•	Meal compatibility is currently heuristic-based
•	Genetic Algorithm typically performs better than Simulated Annealing for this problem

Future Improvements

---

•	Stronger meal realism constraints
•	Better mapping between food categories and meal types
•	Multi-day planning with grocery optimization

Greedy Initialization (initialize.py)

Generates a starting meal plan for the simulated annealing optimizer. Rather than starting from a random or hand-supplied state, this module builds an initial plan that is already reasonably close to the user's nutritional targets, which speeds up SA convergence.

How it works

For each remaining meal slot:

1. Divide the remaining nutrient budget evenly across unfilled slots.
2. Sample a pool of candidate foods at random from the food database.
3. For each candidate, compute the serving multiplier that best hits the per-slot calorie target, then score it by how well it covers the other nutrient targets (protein, fat, carbs, fiber).
4. Apply a soft penalty to candidates whose category has already appeared in the plan, to encourage variety.
5. Pick the lowest-scoring candidate and add it to the plan.
6. Subtract that food's nutrient contribution from the remaining budget before moving on to the next slot.

Parameters

foods_df Cleaned food database (output of preprocess.py).

logged Nutrients already consumed today, e.g. {"energy_kcal": 400, "protein_g": 15, ...}.

targets Full-day nutrient targets, e.g. {"energy_kcal": 2000, "protein_g": 50, ...}.

n_slots Number of remaining meal slots to fill.

penalty_cfg Config dict shared with the SA scorer. Must contain "portion_min" and "portion_max". Optionally contains "variety_penalty" (default 1.2), a score multiplier applied when a candidate's category is already in the plan.

candidate_pool_size How many foods to evaluate per slot. Higher values improve initial plan quality at the cost of speed. 50 is a reasonable default for most dataset sizes.

seed Random seed for reproducibility.

Notes

• The initializer only considers targets (lower-bound nutrients like calories, protein, fiber). Upper limits like sodium and sugar are not penalized here -- the SA scoring function handles those.

• Serving sizes are anchored to the per-slot calorie target. A food providing 200 kcal per 100g in a 500 kcal slot will be assigned a multiplier of 2.5 (250g), clamped to [portion_min, portion_max].

• If the food database is very small, the pool sampling falls back to allowing repeated foods to avoid failure.