On April 1, 2026, the scientific article titled "Recurrence mimicking learning: Eliminating sequential rollouts in offline recurrent reinforcement learning" was published in the journal Neurocomputing (and this is no April Fools' joke), presenting an interesting proposal for enhancing the training processes of recurrent neural networks. Authored by Tomasz Witkowski, M.Sc., Prof. Krzysztof Kania, and Prof. Tomasz Wachowicz, the paper demonstrates how to significantly accelerate the development of decision-making tools used in dynamic decision-making problems.
The article is available directly at the following link: https://www.sciencedirect.com/science/article/pii/S0925231226002043
Abstract:
Recurrent Reinforcement Learning (RRL) is widely used in settings where actions depend on previous decisions, such as dynamic decision-making. However, offline RRL suffers from a major computational drawback: it evaluates trajectories step by step, making training inefficient for long horizons, complex models, and high-dimensional features. To address this, we propose Recurrence Mimicking Learning (RML), an approach that reorders offline RRL rollouts to require only two batched forward passes per epoch, independent of horizon length. RML enumerates all previous actions in a single pass and reconstructs the exact recurrent path through a lightweight selection step. Experiments show that RML preserves the exact final action trajectory of standard offline RRL, allows direct optimization of global rewards, and reduces training computation time to approximately 5% of the conventional approach, while scaling efficiently with both sequence length and action space size.
