Monte-carlo tree search for 3D packing with object orientation

Abstract

In this paper we look at packing problems that naturally arise in container loading. Given a set of 3D iso-oriented objects and a container, the task is to find a packing sequence of the input objects consisting of the ID, location, and orientation that minimizes the space wasted by the packing. Instead of the decision problem, we look at the packing optimization problem, minimizing the total height of a packing. Our solutions uses extreme points and applies Monte-Carlo tree search with policy adaptation, a randomized search technique that has been shown to be effective for solving single-agent games and, more recently, complex traveling salesman and vehicle routing problems. The implementation is considerably simple and conceptually different from mathematical programming branch-and-bound and local search approaches. Nonetheless, the results in solving 2D and 3D packing problems are promising.