A Deep Reinforcement Learning Approach to Packaging Optimization

Loading...
Thumbnail Image
Date
Authors
George, Graeme
Keyword
Packaging Optimization , Artificial Intelligence , Reinforcement Learning
Abstract
Packaging optimization is the process of determining an optimal layout of components in an assembly with the goal of maximizing packing density. Existing computational tools for solving packaging optimization problems are focused in their scope, as they are often tailored to specific problems, and typically require significant effort from the designer to tune relevant parameters. Packaging optimization is also complex due to the multi-modal and non-linear objective functions, which makes the task difficult to solve using existing combinatorial optimization methods. Advancements in computational power and memory have increased the applicability of machine learning to new areas, including design optimization. Reinforcement Learning (RL) is a subset of machine learning which is well-suited to optimization tasks. The use of artificial neural networks (ANNs) within RL is called Deep Reinforcement Learning (DRL). DRL has been successfully applied to a wide array of complex tasks for which traditional computational methods have not been well-suited to. This work proposes a methodology for defining both 2D and 3D packaging optimization as a DRL task. The development of the problem definition is discussed in detail, including the selection of an appropriate DRL algorithm and the tuning of critical algorithm parameters. Several case studies for 2D and 3D packaging optimization are developed to train and evaluate the DRL formulation. Two variations on the reward function are defined to examine the effect feedback frequency on agent performance. High-quality solutions are produced for simple 2D case studies, demonstrating the ability of DRL to solve some packaging optimization problems. Sub-optimal results are produced for the more complex 2D case studies and all the 3D case studies. The formulation is therefore limited in its applicability, but several areas of improvement are identified and discussed. This work provides a proof of concept for the development of a generalized packaging optimization tool.
External DOI