Experiments

🏃‍♂️ Training and Test Set

RS

The sequences are generated by sampling items out of the pre-defined item set randomly. The disvantage is the optimality of a sequence is unknown so that it is difficult to gauge the packing performance with this benchmark.

Cutting Stock

Generating training sequences via cutting stock, by which items in a sequence are created by sequentially "cutting" the bin into items of the pre-defined types, so that the sequence may be perfectly packed and restored back to the bin.

Two variations:

• CUT-1: sorting items based on Z coordinates of their FLBs
• CUT-2: sorting items based on their stacking dependency

💯 Performance

The performance of the packing algorithm is quantitated with

• space utilization (space uti.)
• the total number of items packed (# items)

🦎 Extension to Different 3D-BPP Variants

• Multiple bins:
  • initialize multiple BPP-1 instances matching the total bin number
  • choose the bin in which the item introduces the least drop of the critic value given by the corresponding BPP-1 network
• Item re-orientation:
  • create two feasibility masks for each item, one for each orientation