Bridging Theory and Practice: Robust Deep Reinforcement Learning for IRS-Assisted IoT Under Imperfect CSI

Intelligent reflecting surfaces (IRS) coupled with deep reinforcement learning (DRL) promise significant gains for energy harvesting (EH) Internet-of-things (IoT) networks. However, most of the existing literature assumes perfect channel state information (CSI) while modeling, which is an assumption that is invalidated in real deployments with noisy and outdated information. This work introduces a realistic framework for training DRL agents under channel imperfections characterized by estimation errors, temporal correlation, and quantized feedback. We explore a case study of IRS-assisted EH in smart city environments, which demonstrates practical deployment considerations across diverse IoT applications. Comprehensive experiments were conducted using deep deterministic policy gradient (DDPG), twin delayed DDPG (TD3), proximal policy optimization (PPO), and soft actor-critic (SAC). Results demonstrate a 31% throughput improvement with the introduction of just 20 IRS elements as compared to a non-IRS baseline under perfect CSI when averaged across the four algorithms. Moreover, under moderately imperfect channel conditions, the best-performing algorithm in each CSI scenario exhibits no more than 10% degradation relative to the perfect CSI benchmark. These findings provide actionable deployment guidelines for IRS-assisted EH IoT systems operating under real-world channel conditions.