Abstract

Reconfigurable Intelligent Surfaces (RIS) have emerged as a transformative technology for next-generation wireless networks, enabling programmable control of electromagnetic wave propagation through arrays of passive reflecting elements. Beyond their well-established role in enhancing communication capacity and coverage, RIS panels offer exceptional opportunities for high-precision indoor positioning by creating additional controllable propagation paths that can be exploited for geometric localization. This paper presents a deep reinforcement learning (DRL) framework for joint position estimation and adaptive beam management in multi-RIS indoor environments that leverages time-division RIS operation combined with learned signal processing to achieve accurate user equipment (UE) localization in challenging non-line-of-sight (NLOS) conditions. The proposed architecture comprises three synergistic components: i) a deep neural network (DNN) position estimator that maps time-difference-of-arrival and received signal strength features directly to UE coordinates, implicitly learning to handle NLOS bias and clock synchronization errors from augmented training data; ii) a Deep Q-Network (DQN) that learns an adaptive RIS anchor selection policy to identify and exclude NLOS-corrupted measurements; and iii) a hybrid geometric refinement stage where the DNN-predicted position initializes Levenberg-Marquardt optimization on the DQN-selected anchor subset. The framework is evaluated through extensive Monte Carlo simulations in a representative 60×40m indoor environment with four strategically deployed 256-element RIS panels operating at 3.5GHz with 100MHz bandwidth. Comparative analysis against joint Levenberg-Marquardt estimation, standard TDoA processing, robust TDoA with iteratively reweighted least squares, and K-Nearest Neighbors fingerprinting demonstrates that the proposed DQN-DNN-LM pipeline achieves a median positioning accuracy of 1.41m, representing a 27.4% improvement over the best geometric baseline while providing increasingly significant advantages under high NLOS contamination. Comprehensive parametric studies examining the impact of signal-to-noise ratio, RIS array size, and NLOS probability reveal the fundamental scaling behavior and practical design trade-offs governing DRL-aided RIS positioning systems.