Author: Oluwatobi Owoeye,”A Geometric-Feature-Driven Machine Learning Framework for Predicting Optimal Packing of Irregular Convex Shapes” Handsonlabs Software Academy Initial Paper release
Abstract:
This paper presents a novel machine learning framework for predicting the optimal spatial arrangement of irregular convex shapes a problem inspired by the packing of seasonal decor such as Christmas trees. We generate a dataset of 200 distinct geometric configurations via an iterative simulation-based optimizer, each characterized by 13 engineered features describing spatial distribution, density, angular orientation, and inter-object distances. Using a rigorous data-splitting strategy (train/validation/test/holdout), we compare multiple regression models and identify Gradient Boosting as the best-performing predictor, achieving an R² of 0.999 on training data and generalizing robustly to unseen data (holdout R² = 0.995). Notably, generalization gaps remain small (≤0.014), and no overfitting is detected. The model accurately estimates the normalized convex hull area with a mean absolute error as low as 0.046 on training samples. This work demonstrates that high-level geometric descriptors can be leveraged by ML models to reliably predict packing efficiency, offering a scalable surrogate for costly combinatorial optimization in spatial design and logistics.
Keywords: Geometric feature engineering, packing optimization, convex hull, Gradient Boosting, regression model, generalization analysis, surrogate modeling, spatial machine learning.