Electronic Theses and Dissertations (Masters)

Permanent URI for this collectionhttps://hdl.handle.net/10539/37972

Browse

Filters

20241

Settings

Author: search.filters.author.Garg, ArnavSubject: search.filters.subject.INVENTORY MANAGEMENT

Search Results

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Inventory Management using Artificial Intelligence
    (University of the Witwatersrand, Johannesburg, 2024) Garg, Arnav; Smith, Bevan; Rich, William
    Poor inventory management negatively affects a company’s profits. Too little stock limits potential sales and customer satisfaction while too much stock increases storage costs and potential damage. Company X distributes butter in a multi-echelon supply chain consisting of multiple entities such as a manufacturing plant, distribution centre and retailers before reaching the customer. For every 1% of the demand that is not met, revenue in excess of R18 million is lost per year. This study aims to use machine learning methods (supervised and reinforcement learning) for optimal decision making that maximizes profits. Supervised learning methods (random forest, recurrent neural network and support vector machines) were used to forecast the demand based on historical data. Thereafter, deep reinforcement learning (DRL) was used to train an agent to decide when and how much to order over a period of a year. Various algorithms (PPO and DDPG) and unique reward functions were tested and the performance was compared to a benchmark heuristic that stocks inventory based on a sum of the forecasted demand. The random forest algorithm performed the best at predicting the forecasted demand. The DRL model using a continuous action and state space together with the DDPG algorithm and a reward function based on a combination of the current profit, order fulfilment rate, units available and units unsatisfied performed the best. The DDPG algorithm outperformed the PPO algorithm with the DDPG model being able to provide a 21% increase in net profit over the benchmark heuristic when the production and warehouse facilities of the supply chain were merged. The DRL models were not able to provide a higher order fulfilment rate compared to the benchmark heuristic but they were able to provide better asset utilization by sending full trucks and minimizing the inventory held to maximize the profitability in the supply chain. The results suggest that RL has potential of better handling the stochastic constraints (demand and lead times) in real supply chains to automate the ordering process. It was found that increasing the order fulfilment rate does not necessarily lead to higher profits and the reward function has a significant effect on the net profit which can be further optimized in the future.