GPU-Accelerated Robust MA--RIS Pre-Optimization for Communication and Target Illumination

Tran Le Dung; Ho Viet Duc Luong

cs.ITOct 15, 2026

GPU-Accelerated Robust MA--RIS Pre-Optimization for Communication and Target Illumination

Authors: Tran Le Dung, Ho Viet Duc Luong

Organizations: School of Information and Communications Technology (SoICT), Hanoi University of Science and Technology (HUST), Hanoi 100000, Vietnam

Abstract

Movable antennas (MAs) and reconfigurable intelligent surfaces (RISs) offer strong spatial control for integrated sensing and communication (ISAC), but repeatedly adapting MA positions and RIS phases to each channel realization can be impractical. This letter develops a GPU-accelerated robust pre-optimization framework that jointly selects slow-timescale MA geometry and RIS phases and returns a reusable Pareto set balancing expected worst-user rate and transmit-side worst-target illumination. The illumination objective measures directional transmitted power rather than complete echo-detection or parameter-estimation performance. The proposed Regime-Aware Multiobjective Pre-optimization (RAMP) preserves regime-specific Pareto populations, regulates cross-regime mating using Pareto-survival feedback, and constructs a worst-regime robust archive. Numerical results demonstrate competitive robust Pareto quality against the considered evolutionary baselines while avoiding online physical re-optimization.

Figures & tables

Fig. 1: MA–RIS downlink considered in this letter. The MA geometry $\mathbf{P}$ and RIS phase vector $\boldsymbol{\theta}$ are optimized offline and fixed during operation; only communication beamformers are recomputed online.

Approach	Per-regime	Transfer signal	Robust set
MGRAP pre-opt. [ 10 ]	Yes	Scalar	No
Pooled NSGA-II	No	None	No
Independent NSGA-II	Yes	None	Yes
RAMP	Yes	Pareto survival	Yes

TABLE I: Positioning against representative pre-optimization and search strategies.

Parameter	Value
Carrier; $M,K,Q$ ; RIS	$28$ GHz; $16,4,3$ ; $8\times 8$ ( $L=64$ )
MA/RIS geometry	$4\times 4$ cells; cell pitch $\lambda$ ; cell extent $\lambda/2$ ; $d_{\min}=\lambda/2$ ; RIS $(30,40,10)$ m, spacing $\lambda/2$
Users / target directions	users at $50$ – $200$ m in a $60^{\circ}$ sector; far-field target directions with $20^{\circ}/20^{\circ}$ offset/spread
Channel	$61.4+22\log_{10}d$ dB; no shadowing; $4$ paths, $120^{\circ}$ spread; Rician $6$ dB
Power/noise/EVM/RIS	$30$ dBm; $180$ kHz; $9$ dB; $\kappa=0.05$ ; unit-modulus RIS; $\eta=1$
Optimizer	$N_{\rm p}=200$ , $G=300$ , $B=32$ ; $r_{0}=0.3$ , $[0.05,0.5]$ , $\alpha=0.05$ , $\beta=0.1$

TABLE II: Main simulation and optimization settings.

Fig. 2: Unified worst-regime fronts; the second axis is transmit target illumination. HV reference:

(0,-30\,\mathrm{dB})

\text{HV}_{\text{RAMP}}=186.05

Method	HV	IGD	$95\%$ CI
RAMP (adaptive)	$\mathbf{182.59\pm 19.1}$	$\mathbf{0.018}$	–
No transfer ( $r=0$ )	$178.23\pm 20.0$	$0.066$	$[3.17,5.55]$
MO-DE	$178.16\pm 19.9$	$0.128$	$[2.79,6.07]$
Fixed transfer ( $r=0.3$ )	$176.45\pm 19.8$	$0.150$	$[4.71,7.63]$
Pooled	$174.86\pm 20.3$	$0.758$	$[5.56,9.90]$
MOPSO	$167.71\pm 17.2$	$1.434$	$[12.34,17.42]$

TABLE III: Robust-front quality over

15

paired seeds. CI is for

\Delta

HV = RAMP

-

method.

Fig. 3: Mean best-so-far robust HV over the $15$ paired seeds.

Matched GPU speedup: $(4.12\pm 0.20)\times$ total; $3.09\times$ search-only.
Method	Time (s)	Memory (MB)	Eval./s
RAMP	$80.0$	$3822$	$1845$
No transfer	$76.6$	$3822$	$1941$
MO-DE	$48.8$	$318$	$2805$
Pooled	$56.3$	$2239$	$2726$
MOPSO	$45.8$	$366$	$1587$

TABLE IV: Measured computational cost at the default setting.

Fig. 4: Per-realization HV CDF using $200$ offline and $30$ online-reference realizations, with $75$ generations per online run. The shaded region denotes the $10$ – $90\%$ interval of the online reference.

Multi-Objective Optimization

Multi-objective optimization studies how to find and choose solutions that balance competing goals, constraints, and preferences.

See all 266 papers

A Particle-Swarm-Assisted Gradient Meta-Learning Algorithm for Joint Transmit Precoding and STAR-RIS Coefficient Optimization
Sep 24, 2026Kang ZhouReconfigurable Intelligent Surfaces Learning to Optimize
School of Artificial Intelligence, Mianyang City College, Mianyang 621000, China
Trustworthy AI-Driven Dynamic Hybrid RIS: Joint Optimization and Reward Poisoning-Resilient Control in Cognitive MISO Networks
Mar 27, 2026Deemah H. Tashman, Soumaya CherkaouiWireless Communications Reconfigurable Intelligent Surfaces
LINCS Laboratory, Department of Computer and Software Engineering, Polytechnique Montréal, Montréal, QC, Canada, H3T 1J4
Active Sensing for RIS-Aided Tracking and Power Control: A Hybrid Neuroevolution and Supervised Learning Approach
Jun 30, 2026George Stamatelis, Hui Chen, Henk Henk Wymeersch +1Reconfigurable Intelligent Surfaces Active Sensing
Department of Informatics and Telecommunications, National and Kapodistrian University of Athens, Panepistimiopolis Ilissia, 16122 Athens, Greece · Department of Electrical Engineering, Chalmers University of Technology, 412 58 Gothenburg, Sweden

Sep 24, 2026J/K move · Enter open · S save

A Particle-Swarm-Assisted Gradient Meta-Learning Algorithm for Joint Transmit Precoding and STAR-RIS Coefficient Optimization

Kang Zhou

School of Artificial Intelligence, Mianyang City College, Mianyang 621000, China

Reconfigurable Intelligent Surfaces Learning to Optimize Wireless Communications Wireless Resource Allocation

Open paper AskPDF arXiv

This paper investigates the joint optimization of the transmit precoder and the transmission/reflection coefficients of a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) to maximize the weighted sum rate (WSR) in a multi-user downlink. We propose a particle-swarm-assisted gradient meta-learning (PSA-GML) algorithm for this non-convex problem. The original problem is first equivalently transformed via an amplitude-split parameterization and a collapsed precoder representation, which automatically satisfy the energy-conservation constraint and reduce the search dimension. Particle swarm optimization (PSO) then performs a global search over the STAR-RIS coefficients to yield a high-quality, initialization-robust warm start, with the transmit precoder obtained in closed form. Departing from conventional alternating optimization (AO), a coordinate-wise long short-term memory (LSTM) meta-optimizer trained by first-order gradient meta-learning further refines the coefficients and precoder jointly, learning per-coordinate adaptive update rules from data. The meta-optimizer is trained offline and applied to unseen channels without further adaptation. Numerical results show that PSA-GML attains an 11.06 bits/s/Hz WSR at 10 dB with N=32 elements and K=4 users, exceeding AO by 13.1% (and by 6.2% even with multiple random restarts) and the random-phase scheme by 35.1%. In the interference-limited regime it reaches 83.9% of the hand-designed Adam refinement without manual hyper-parameter tuning, and it transfers zero-shot across regimes, indicating that the learned update rule captures the intrinsic WSR landscape structure.

Abstract

Multi-Objective Optimization

Explore similar work

A Particle-Swarm-Assisted Gradient Meta-Learning Algorithm for Joint Transmit Precoding and STAR-RIS Coefficient Optimization

Trustworthy AI-Driven Dynamic Hybrid RIS: Joint Optimization and Reward Poisoning-Resilient Control in Cognitive MISO Networks

Active Sensing for RIS-Aided Tracking and Power Control: A Hybrid Neuroevolution and Supervised Learning Approach