Symmetric-PCGrad: A Conflict-Aware Approach for Multi-Task Learning in RIS-Enabled Networks

Reconfigurable Intelligent Surfaces (RIS) provide a fivefold increase in signal power gain compared to traditional approaches, while being easy to deploy and offering cost-effective solutions. A key factor in RIS performance is the intelligent activation of element groups—for instance, enabling only selected elements for reflection while keeping the rest in absorption mode to minimize unnecessary power consumption. Motivated by the need for efficient resource utilization, this study proposes a multi-task deep learning framework based on Symmetric-Projecting Conflicting Gradients (Symmetric-PCGrad). The framework jointly performs two tasks: (i) classifying the active group of RIS elements responsible for producing high signal power at specific azimuth and elevation angles, and (ii) estimating the received signal power at the user end. This dual-task approach is designed to identify the minimal group of elements required to meet a target quality of service—an aspect largely overlooked in existing RIS-assisted wireless communication research. To ensure fair evaluation, we applied the same multi-task learning (MTL) architecture using three strategies: standard MTL (no gradient projection), MTL with Projecting Conflicting Gradients (PCGrad), and the proposed Symmetric-PCGrad, all under identical hyperparameter settings. Experimental results demonstrate that our method outperforms both baseline techniques in terms of individual task performance and overall multi-task effectiveness.