Multi-stage predictive engine for low-fidelity sensors

A low-fidelity sensor technology is developed by MOOS which operates in a setting which demands very high predictive accuracy. These sensors are paired with software to gain insight into inventory positions, allowing companies to optimise their operations. From an AI perspective, the effort is centred around the development of a multi-stage predictive algorithm which is able to achieve high predictive accuracy in settings where sample sizes are relatively small, while at the same time flexible non-parametric (e.g. Neural Networks) or ensemble methods (e.g. Random Forests) are required to account for a high degree of complexity in the data. This multi-stage approach is useful in scenarios where the main predictive objective can be broken down into a set of “simpler” predictive targets. The set of first-stage algorithms produce a number of predictions which are used as inputs for the second-stage predictive algorithm. In this sense, the first-stage models deliver a form of controlled and supervised dimensionality reduction step which turns raw data into relevant indicators for the final predictive model. This dimensionality reduction step is however more structured than that achieved by standard clustering techniques or PCA. In small samples, the multi-stage algorithm achieves better results than a standard single-stage non-parametric or ensemble algorithm by imposing structure into the predictive problem and improving on the classic bias-variance trade-off. In specific settings, the multi-stage algorithm is less demanding in terms of the volume of data that is required to achieve high predictive accuracy. Our end result is a plug-and-play AI engine that delivers high accuracy performance for predicting a wide range of product sizes, weights, and shapes in a retail shelf. Regardless of sensor shape of size and irrespective of how products are stored or stacked.