Predictive packaging system

A method implemented by a computer system that hosts a predictive model, the method comprising:

• receiving first item data from a workstation, the first item data comprising a description of an item, the workstation configured to facilitate packaging of the item;
• generating, based at least in part on the first item data, an input to the predictive model, the predictive model trained based at least in part on batches of training data to generate outputs about package decisions, the training data excluding the first item data and including second item data, packaging data, and damage data associated with a different item;
• generating, based at least in part on an output of the predictive model, a package decision indicating a package type associated with the packaging of the item;
• and
• sending, to the workstation, the package decision, wherein the predictive model is iteratively trained by using batches of training data and a loss function, wherein the training comprises:
• receiving a batch of training data associated with packaged items, the batch comprising item descriptions and item images of the packaged items;
• determining actual package types and observed damages associated with the packaged items;
• generating information based at least in part on the item description, the item images, the package types, and the observed damages;
• and
• iteratively inputting the information to the predictive model to generate predicted package decisions, update the loss function, and update parameters of predictive model based on the loss functionswherein the loss function is updated based at least in part on a training output of the predictive model indicating a predicted package type and a predicted damage for using the predicted package type, a first comparison of the predicted package type to an actual package type, and a second comparison of the predicted damage to an observed damage due to using the actual package type.

One or more non-transitory computer readable media storing instructions, that upon execution on a computer system hosting a predictive model, cause the computer system to perform operations comprising:

• receiving first item data from a workstation, the first item data comprising a description of an item, the workstation configured to facilitate packaging of the item;
• generating, based at least in part on the first item data, an input to the predictive model, the predictive model trained based at least in part on damage data associated with packaging items to generate outputs about package decisions;
• generating, based at least in part on an output of the predictive model, a package decision indicating a package type associated with the packaging of the item;
• and
• sending, to the workstation, the package decision, the workstation configured to package the item in a package corresponding to the package type,
• wherein the predictive model comprises at least one of a deep neural network, a convolutional neural network, a recursive neural network, or a reinforcement learning model and is iteratively trained by using batches of training data and a loss function, wherein the training comprises:
• receiving a batch of training data associated with packaged items, the batch comprising item descriptions and item images of the packaged items;
• determining actual package types and observed damages associated with the packaged items;
• generating information based at least in part on the item description, the item images, the package types, and the observed damages;
• and
• iteratively inputting the information to the predictive model to generate predicted package decisions, update the loss function, and update parameters of predictive model based on the loss functions wherein the loss function is updated based at least in part on a training output of the predictive model indicating a predicted package type and a predicted damage for using the predicted package type, a first comparison of the predicted package type to an actual package type, and a second comparison of the predicted damage to an observed damage due to using the actual package type.