Smart nesting: estimating geometrical compatibility in the nesting problem using graph neural networks

Abstract Reducing material waste and computation time are primary objectives in cutting packing problems (C &P). A solution to the C &P problem consists of many steps, including grouping items be nested arrangement grouped on a large object. Current algorithms use meta-heuristics solve directly without explicitly addressing problem. In this paper, we propose new pipeline for nesting that starts with then arranging them objects. To end, introduce motivate concept, namely Geometrical Compatibility Index (GCI). Items higher GCI should clustered together. Since no labels exist GCIs, model GCIs as bidirectional weighted edges graph call geometrical relationship (GRG). We novel reinforcement-learning-based framework, which two neural networks trained an actor-critic-like fashion learn GCIs. Then, group into clusters, GRG capacitated vehicle routing it using meta-heuristics. Experiments conducted private dataset regularly irregularly shaped show proposed algorithm can achieve significant reduction (30% 48%) compared open-source software while attaining similar trim loss regular threefold improvement irregular items.