Meet over 65 universities on 27 & 28 April > REGISTER NOW
Anglia Ruskin University ARU Featured PhD Programmes
University of Reading Featured PhD Programmes

Supply chain Robotic Process Automation through Machine Learning

Faculty of Engineering and Informatics

This project is no longer listed on and may not be available.

Click here to search for PhD studentship opportunities
Dr Amr Abdullatif , Dr Sohag Kabir , Dr Ibrahim Ghafir Applications accepted all year round Self-Funded PhD Students Only

About the Project

In Supply Chain optimization, understanding which item is worth keeping in stock, considering several factors ranging from continuous features, like demand pattern clustered in different time buckets, to categorical features like supplier name, and internal tagging like obsolete or special certification needed, is a key task for the planner. All those features should be associated to a label that will define synthetically the importance of the item for the sales process and its risk in terms of unsold inventory, in few words we want to answer the question “shall I keep this item in stock or not?”. The answer to this question is historically a mix of analytic skills and expertise from a dedicated planner who is using static statistical models based on average consumption corrected by a z-score proportional to process variability to ensure a pre-determined service level.

The required machine learning model should compose of 2 processes:

· Feature engineering of the demand patterns by building a clustering process having as its goals to cluster demand patterns chunks into c clusters, where chunks with high membership to the same cluster represent similar temporal patterns, and at the same time to measure the outlierness degree of each chunk and consequently to measure the density of outliers. This to track the changes in the incoming demand patterns. The goal of this first process is to build a clustering model that is able to adapt to the changes in the demand patterns, by implementing a continuous learning that exploits the input chunks as they arrive. Intrinsic to this clustering model is a measure of outlierness that provides information about the goodness of fit of each input chunk to the clustering model.

· The second process is applying a supervised learning algorithm that is used to mimic the expert opinion of the planner, and decide whether we should set an item for make to stock (MTS) or make to order (MTO). This is called MTS/MTO classification.


· Integrate the domain knowledge and machine leaning to create a new classification and level setting process, leveraging years of demand data and new statistical indicators for demand patterns.

· The validation phase should be reduced at each iteration as ML model can be re-trained to incorporate past validations, increasing efficiency and performances.

· Level setting problem (on top of the deployed classification model) can be addressed by benchmark Machine Learning methods (e.g. Reinforcement Learning, Montecarlo simulations and traditional statistical methodologies). Regarding Reinforcement learning and Montecarlo you have to establish punishments for letting a particular item run out of stock and also punish the model for stock with too higher value for too long. For rewards, the focus was on ordering items within a safe window before the demand.


1. A. Abdullatif, F. Masulli, and S. Rovetta. Tracking time evolving data streams for short-term traffic forecasting. Data Science and Engineering, 2(3):210–223, Sep 2017.

2. A. Abdullatif, F. Masulli, and S. Rovetta. Clustering of nonstationary data streams: A survey of fuzzy partitional methods. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 8(4):e1258, 2018.

3. A. Abdullatif, F. Masulli, S. Rovetta, and A. Cabri. Graded possibilistic clustering of non-stationary data streams. In A. Petrosino, V. Loia, and W. Pedrycz, editors, Fuzzy Logic and Soft Computing Applications, pages 139–150, Cham, 2017. Springer International Publishing.
Search Suggestions

Search Suggestions

Based on your current searches we recommend the following search filters.

FindAPhD. Copyright 2005-2021
All rights reserved.