Mitigating exposure bias for reliable ANN-based control: The enhanced scheduled sampling framework

The design of Artificial Neural Networks (ANNs) as controllers for industrial processes offers significant potential but is often hindered by the challenge of exposure bias. Classic training methods, such as Batch Learning (BL), can suffer from it when trained controllers are deployed in closed-loop systems. This paper introduces the Enhanced Scheduled Sampling (ESS) framework, a novel and comprehensive training methodology designed to mitigate these limitations and produce reliable ANN-based controllers. Based on the Scheduled Sampling curriculum learning strategy, the ESS framework integrates an online simulation environment with key enhancements such as Weighted ANN Actuations to blend controller outputs, a sequence-based control error term and a controlled dynamic learning rate adjustment strategy. The framework is validated against the baseline learning approach on a conventional PID control scenario acting on First-Order Plus Dead-Time processes as well as on a non-linear Continuous Stirred-Tank Reactor exhibiting second-order dynamics. The PID benchmark provides an industrially relevant reference, enabling the proposed ANN-based controller to be compared with the established standard for process-level feedback control. The results demonstrate that controllers trained with ESS consistently outperform their BL counterparts in both control performance and robustness to parametric variations and training initialisation. By effectively mitigating exposure bias, the ESS framework represents a significant step towards developing dependable, high-performance ANN controllers ready for real-world industrial deployment or to advance research in transfer learning for control. The open-source implementation is shared to encourage further research.