Learning heuristics for a supply planner

Abstract

Supply planning is an NP-Hard problem that is often tackled when dealing with supply chain management. It is a problem with many variations but the core idea is to create a plan that resolves as much demand as possible.
There are different approaches in use to solve a planning problem, one of which is by using heuristics. Heuristics are used to get to a relatively good solution within some feasible amount of time compared to trying to get to an optimal solution. Expertise of the problem is often required to come up with good heuristics or to improve existing ones. When such an expert is not readily available alternatives need to be considered, one of which is machine learning.
The goal of this thesis is to create a machine learning model that can improve the heuristics used in an existing planning algorithm at Outperform. This heuristic decides upon a sequence in which products are included in the planning algorithm. Using a different sequence can lead to vastly different results for the algorithm both in quality and time required. Our focus will be on reducing the amount of time required for the algorithm by learning a heuristic that can provide better sequences. 
To achieve this goal, two models are created and evaluated. These models use imitation learning to replicate the sequences that were the fastest. Due to the lack of an expert to provide the fastest sequences, an oracle is created that attempts to search for a sequence as fast as possible within some feasible amount of time. The trained models are evaluated on several supply chains provided by Outperform.