By Daniel Glass
Originally Posted on LinkedIn
We have talked a lot about the fundamental concepts of ML and how relevant it is to so many applications. BUT, what is the limit here? I want to give an example of one limit with our warehouse task time prediction model. This limit also applies to other applications of ML, including LLMs -- you can think on your own about how LLMs might be affected by this.
Let's say that we have trained our task time prediction model on every task ever performed in our warehouse. Our model adequately predicts how long tasks should take given certain characteristics (travel distance, speed, activity type, associate experience, etc.). We are able to use this to determine which associates need training, to estimate how much labor is needed to meet daily objectives, which associates are going above and beyond, and so on.
BUT, what if an eagle flew in through one of the opened doors, snatches a case that a user is picking, flies around the warehouse for a minute or two, and drops the case 100 feet away from the picker before the eagle flies out the way it came. The actual amount of time this task took vs what our model predicts is going to be very different. We do not have an "Eagle took my case and flew around with it" input characteristic for our model to use during prediction. This is a silly example, but it illustrates that our models will only be as good as the data we give them. It is very difficult to know all of the things that could influence reality and even more difficult to neatly bundle it all up as input for our models.
Stay tuned for our last post in this series in a few days!
If you haven't yet Check out how Cellaware is changing the way AI/ML is used in warehousing operations by scheduling your free demo of ChatWMS.
Comments