ReEvo evolves crossover, mutation, and local-search operators offline for a memetic algorithm that secures fluid-antenna port selection, beating GA and two LLM-GA baselines in secure sum-rate

A ReEvo-designed memetic optimizer in which an LLM evolves three complementary operators offline—crossover, mutation, and local search—and embeds them in a memetic search framework that performs no LLM queries online. Prior LLM-GAs evolve only one operator: Ref. [13] designs only a crossover with no explicit local-search stage and no secrecy objective, while Ref. [14] designs only a one-port mutation targeting nonsecure max-min SINR; this work covers all discrete search stages. A table contrasts which LLM-designed operators (mutation, crossover, local search) appear in [13], [14], and the proposed method, and the text states the three operator classes are trained separately and frozen for online use.

Secure WMMSE beamforming supplies the secrecy-driven fitness, so operator selection is driven directly by the secrecy objective rather than a surrogate port score. Ref. [13] uses power minimization and Ref. [14] uses nonsecure multiuser performance, whereas here the meta-fitness is the best secure sum-rate attained when an inner GA-based search uses that operator on a fixed training channel set. The meta-fitness definition (16) is given and stated to include secure WMMSE optimization; the beamforming procedure is a safeguarded WMMSE with backtracking halving (at most 20 trials), an acceptance tolerance, and termination after 30 iterations or improvement below a threshold.

At equal generation counts, the proposed algorithm attains the highest mean secure sum-rate at every tested transmit power, with both LLM-GA baselines consistently lying between GA and the proposed method. Relative to conventional GA and to LLM-GAs that design only crossover or only mutation, evolving all three operators plus systematic local refinement yields higher values on the secrecy objective. Simulations report secure sum-rate and gains over GA, LLM-GA-crossover [13], and LLM-GA-mutation [14] at power points including 1 W and 5 W; error bars show one sample standard deviation of the 20 per-channel total rates, and error bars visibly overlap at some operating points.

A system model and problem formulation for secure port selection: exactly a fixed number of distinct ports are activated on a uniform grid over a rectangular FAS aperture, and the port selection matrix and beamforming matrix are jointly designed to maximize the users' secrecy sum-rate. The joint design of discrete port selection and continuous beamforming under passive eavesdropping is unified as a nonlinear mixed-integer nonconvex program P1, with the explicit observation that selecting a port changes the sampled channels of every user and of the eavesdropper simultaneously. Feasibility constraints (each column selects one physical port; selected ports must be distinct), the transmit-power constraint, single-user-decoding SINR expressions, and the per-user secrecy rate built on the wiretap construction (Eq. 10) are given, and P1 is identified as a nonlinear mixed-integer nonconvex program.

The result applies to multiuser downlink FAS with single-antenna users and one passive single-antenna eavesdropper under quasi-static channels, and assumes the base station has perfect instantaneous CSI for every link (eavesdropper CSI may have been acquired while that terminal was previously active), so it serves as an upper-bound benchmark. Within this setting, the pipeline of evolving three operators offline and executing them frozen online suits deployments that need secure port selection without online LLM calls, such as military, satellite, and IoT networks. The authors leave extension to imperfect or delayed CSI and to more general multiuser and multi-antenna scenarios as future work.

Several quantities appear as placeholders in the text (for example aperture dimensions, port counts, user counts, and the specific rates and gain percentages at the power points), so those numbers cannot be verified from the loaded text; the curve details of Fig. 2 and Fig. 3 are also not fully expanded in the prose. Error bars visibly overlap at some operating points, so the statistical robustness of differences among baselines remains for the reader to judge against the original figures. In addition, local search evaluates extra one-swap neighbors, so the convergence curves do not establish faster convergence in runtime or evaluation count, and the monotone objective-value convergence stated for the beamforming procedure is not a beamformer or stationary-point convergence guarantee.