Container Yard Pre-Marshalling Optimization

In a container yard or a steel stockyard, the box you need next is often buried under boxes you need later. Digging it out costs a rehandle every time. This computes, in advance, the shortest sequence of moves that re-sorts the stacks so those rehandles never happen. The industry calls the re-sorting pre-marshalling; we solve it with reinforcement learning.

Where this applies

Anywhere goods are stacked several tiers high and taken out one at a time, the same problem shows up, whatever the size of the site.

Shipping containers stacked several tiers high in rows across a port yard — Container terminal yard. When the stacking order does not match the retrieval order, whatever sits on top has to move first.

Steel wire coils stacked in tiers in an outdoor stockyard with an overhead crane above them — Steel stockyard. Coils and plates stacked in tiers hit the same problem.

Cardboard boxes stacked high in tiers inside a small warehouse — Warehouse. Small sites hit it too, wherever goods go in stacked and come out one at a time.

Photographs illustrate the setting. They are not Hannune project sites.

The problem it solves

Container terminals and steel stockyards store goods in stacks, several tiers high. When the order they are stacked in does not match the order they will leave in, an item that leaves early ends up underneath one that leaves late.

Getting to it means lifting the item on top and putting it somewhere else first. That move is a rehandle, and rehandles are pure cost: crane time spent moving something nobody asked for. Terminals already shift containers during quiet hours to avoid them, which is where the term housekeeping comes from.

This program takes the current state of a bay and computes the move sequence that ends with every stack in retrieval order, using as few moves as it can find. It runs before retrieval starts, not during it.

The method is a re-implementation based on two granted patents held by the inventor, Taeho Kim, founder of Hannune Inc. A neural network scores how good a state is, and a search procedure follows that score to pick the move order.

Measured results

These numbers come from running this program. We have not verified that the solutions are optimal, and we have not yet compared them against the published benchmark instances. What is shown is the behaviour of a model trained to reduce rehandles and move counts.

Problem	Size	Worst-case layout (fixed evaluation state)	20 random layouts (beam search)
Container terminal	8 stacks x 5 tiers, 32 containers, all with distinct retrieval order	34 moves	All 20 solved, 24.1 moves on average
Steel stockyard	8 stacks x 4 tiers, 24 plates	15 moves	All 20 solved, 10.7 moves on average

Measured over 20 random layouts.

Demo

Pick a problem (container yard or steel stockyard) and a layout, then step through the computed move sequence one move at a time.

Problem Case Search

In place Blocking (an earlier-leaving block sits underneath)

Speed 1x

Patents behind it

This program is a re-implementation based on the methods in two granted patents, inventor Taeho Kim.

KR102325608B1

Granted 2021-11-15

Method for generating training data for a neural network applying deep reinforcement learning to pre-marshalling at container terminals

KR102853422B1

Granted 2025-09-04

Method and apparatus for optimal path search using an artificial neural network

How we work on it

Pricing is not published. These are the ways we take this on.

Custom build

Built against your terminal operating system or yard management data.

Proof of concept

Run it on your own yard data first and see what the move count does.

Joint research

Participation in port and logistics research projects or joint ventures.

Enquiries: contact@hannune.ai