Designing a double deep reinforcement learning selection tool for resilient demand prediction

Abstract

Supply chain optimization has attracted many scientific studies for decades. This topic continues to evolve due to the increasing worldwide commercial interactions and the recent progress of artificial intelligence. Artificial intelligence models have emerged as competitive forecasting methods for several years and consumer demands of goods remain a hot topic in supply chain optimisation. However, the process of selecting an appropriate forecasting solution becomes a daunting task, particularly for non-experts because they exhibit a strong dependency on databases. Even specialists face potential errors, considering that each model aligns better with a specific set of problems. This complexity arises due to the distinctive features inherent in each dataset, such as high seasonality, substantial bias, noise, volatility, intermittence, and more. This research endeavors to construct a double deep reinforcement learning model, enabling the automatic selection of a forecasting model from the forecasting committee at the time of prediction. The set of forecasting methods available to be selected in this research holds various neural network based forecasting approaches which have been proven to perform well each one in different dataset configurations; such as feed forward neural networks, recurrent neural networks, convolutional neural networks, boosting ensemble neural networks and stacking ensemble neural networks. To assess the model's performance, an empirical study is conducted using two distinct datasets selected for their diversity in order to ensure robustness to variations of the overall solution approach. The results of experiments demonstrate the resilience of the proposed approach when compared to the state-of-the-art methods.