03 Dec AI in WMS: algorithms, machine learning and deep learning

A year ago I wrote that AI was barely visible in WMS. That has changed. AI is moving quickly toward the warehouse floor, although the gap between marketing and reality remains large. It helps to distinguish three kinds of AI: algorithms, machine learning and deep learning.

1. Algorithms: decision rules and optimization

Algorithms have been part of WMS software for years. They compute with predefined logic, and they come in two kinds.

Decision rules follow fixed logic that you adjust through parameters. A WMS uses them, for example, to:

• put new stock in the first free location in the right zone
• allocate available stock to orders on a FIFO basis
• release orders in a fixed sequence

Advantage: easy to understand and predictable.
Disadvantage: parameters have to be maintained by hand whenever the assortment or demand changes.

Optimization calculates the best solution for the current situation, out of all possible combinations. Think of:

• the best location for each item based on current velocity
• the shortest route past all pick locations in a pick list
• a schedule that accounts for dock capacity, staff and delivery times

Advantage: adapts, because it recalculates for every new situation, and the outcome can be explained.
Disadvantage: requires good data and a good model of the warehouse.

This is the domain of operations research and the foundation of a Warehouse Optimization System.

2. Machine learning: software that recognizes patterns

Machine learning learns from historical data and adjusts control automatically. We still see little of this in WMS packages.

Examples:

• Predicting which items are about to move faster
• Deciding smartly when replenishment is needed
• Predicting busy zones and preventing congestion

Advantage: takes management work off your hands and adapts to circumstances.
Disadvantage: sensitive to messy or incomplete data. Machine learning also looks mainly backward and misses new developments.

3. Deep learning: powerful models that process complex signals

Deep learning goes further and can interpret and generate text, images, speech, movement and large data streams.

Examples:

• Warehouse tasks carried out by chatbots and agents
• Automated quality inspection with cameras
• Robots that pick individual items from totes or boxes

Advantage: takes over work that until now required human perception, such as seeing, reading and grasping.
Disadvantage: decisions are hard to explain, which can lead to distrust.

The strength lies in the combination

The three kinds reinforce each other. Predictive models estimate how demand will develop, and optimization calculates what to do with that prediction. In slotting, for example, a forecast of item velocity determines which items will soon be fast movers, and an optimization model then calculates the best layout within the available space and zones. Deep learning adds perception, such as cameras that count stock or robots that grasp. A Warehouse Optimization System brings forecasting and optimization together, on top of execution in the WMS.

AI will change warehouse control drastically, but only with a solid foundation: good data, predictable processes and active management. Because AI has only recently entered WMS, its functions are still very much in development. Check carefully what vendors actually support today, what is still in the pilot phase and what is on the roadmap.

What do you think: where will AI make the difference in warehouses?