Skip to content

Efficiency in DRL Forecasting

Home / Open Questions / Efficiency in DRL Forecasting

Background: Deep reinforcement learning (DRL) frameworks for forecasting often rely on iterative state estimation and may face computational scalability issues in high-dimensional multivariate settings.

Question / Future Work: The absolute training cost of DRL-based state-aware forecasting remains high for large-scale, high-dimensional multivariate systems, creating a significant bottleneck for real-time or resource-constrained applications. Future work is required to optimize training efficiency and extend these mechanisms to broader classes of non-Markovian or complex state transitions.

Why It Matters: Computational efficiency is a critical barrier to deploying DRL-based sequential modeling in real-time, high-dimensional industrial applications where both rapid adaptation and forecasting are required.

Evidence: While relatively long training times can be acceptable in many practical applications, its absolute training cost remains high when N becomes large, which may limit its practicality in time-sensitive or resource-constrained large-scale settings.

Metadata & Links