Adaptive Beamforming for Low-Power IoT Networks: A Reinforcement Learning Approach

This preprint has not been through journal peer review. Any signed-in platform member may review it, and every review is published under the reviewer's real name, affiliation and ORCID — accountable to write, open for anyone to read. Reviews are the reviewers' own opinions; they are not an editorial decision by Technology Fist.

I read this carefully against the reinforcement-learning beamforming literature and the core contribution holds up: the reward shaping in Section 4 is a genuine improvement over the fixed-penalty formulations most of this field still uses, and the ablation in Table 3 isolates it convincingly. Two things I would want addressed before I would cite this as settled. First, the convergence claim in Section 5.2 rests on twenty runs on a single channel model; with the variance shown in Figure 6 that is thin, and I would want either more seeds or a confidence interval rather than the mean line alone. Second, the energy figures are simulated throughout, and the paper occasionally slips into language ("measured consumption") that implies hardware. That should be tightened, because a reader skimming the abstract would come away believing this was tested on a device. The code and the channel traces are available, which made checking the second point straightforward - credit for that.

A focused, well-scoped contribution. I checked the derivation in Appendix A line by line and it is correct; the step from equation 12 to 13 drops a constant factor that does not affect the result but might briefly confuse a reader reproducing it, so a footnote there would help. The comparison baselines are the right ones and are configured fairly - the authors have used the reference implementations rather than reimplementing competitors, which is where this kind of comparison usually goes wrong. The claimed 18% improvement is consistent with what I would expect from this class of method, so nothing here strikes me as overstated. My only substantive suggestion is that the related-work section under-cites the pre-2020 adaptive-array literature, which solved a version of this problem with very different tools. Situating the contribution against that work would strengthen rather than weaken it.