Multi-Output Regression Indoor Localization Algorithm Based on Hybrid Grey Wolf Particle Swarm Optimization

Round 1

Reviewer 1 Report

The paper concerns the application of multi-output regression to perform indoor localization exploiting WiFi CSI. The main idea is to use a (a slightly modified version of ) Grey Wolf Optimization enhanced with Particle Swarm Optimization to learn the best set of parameters in multi-output SVR. Although the overall idea of not using multiple SVR seems interesting, the paper consists mainly of an application of existing technologies, with some changes in the aforementioned optimization algorithm. Thus, the overall novelty is limited, especially considering that the innovations introduced in the optimization procedure are not properly tested and verified in the evaluation part. Specifically, PSOGWO- MSVR is not tested without both the grey wolf position update strategy with adaptive learning factors and the Tent map initialization using the random variable.

A set of detailed major comments follows:
- The paper does not have a clear structure. Specifically, the related work section is more dedicated at describing technical contents (i.e., material and methods) rather than (more or less) related positioning solutions based on CSI.
- From the title of section 3 it seems that the authors propose Enhanced Hybrid Grey Wolf Particle Swarm Optimization for MSVR Localization. This appears to be an overstatement, as the authors just introduced some adjustment to the optimization algorithm. What is the focus of the paper? Testing MSVR for positioning? Improving the optimization procedure? What is the main contribution? It is not easy to get an answer to these questions by reading the paper.
- In section 4, it is not clear if test reference points are removed completely from the training set (which then becomes of size 35) or not. Moreover, it seems that the scenario is in LOS: this should be specified. Another issue is with what is written in lines 258-267. First, it seems that the choice of relying on amplitude only is done after the evaluation of the performance (which would be wrong). Moreover, there are several papers showing that phase information covers a fundamental role. Concluding that the system works, moreover with just amplitude, considering 35/50 reference points each observed for 20 sec (reduced temporal dynamics) might be quit a bold statement. It follows that also the visual inspection of 3 out of 35/50 reference points is poorly significant. Finally, the noise reduction part appears not well-founded. First, none of the many noise reduction and preprocessing strategies, a very fundamental and critical step for WiFi CSI-bases sensing tasks, developed in the last 10 years of CSI investigation have been considered. Second, the authors propose a not-well explained modification of DBSCAN (A-DBSCAN) as to ignore the parameters it would normally require, yet without considering other clustering approaches (e.g., OPTICS does not require to specify the neighborhood radius).
- Moving to section 5, the first concern relates with the analysis of dimensionality reduction strategies. Besides the facts that other papers investigated this aspect and that in section 4 no explanation are given about the training and evaluation procedures used for the autoencoder employed by the authors, there are not information about hyperparameter tuning approaches for PCA and KPCA. How many principal components have been kept? How much variance has been explained? Why such choices? An out of the shelf usage of such approaches might result in poor performance, biasing towards choosing the autoencoder. In the remainder of the section, the issues are that MSVR optimized without the two enhancements proposed by the authors is not tested (thus, it is not possible to verify the claimed usefulness); no ablation studies are conducted (i.e., it is not possible to weight independently the usefulness of the novel optimization procedure, the usage of the autoencoder, and the usage of A-DBSCAN); no details are given about how the baselines have been obtained (preprocessing of the data? Tuning? …); and, more recent and diverse baselines for CSI positioning should be considered (e.g., look at the works from Christoph Studer, or those using TRSS, etc.)

Minor comments follows:
- lines 30-33 lack proper references. If it is a subjective view/claim, it should be properly supported by evidence.
- comments in lines 65-73 lack support and some are not fully correct.
- MIMO functioning should be introduced
- the usage of intel 5300 NIC for the experiments should be stated early, in order to justify section 2.1
- to the best of the reviewer knowledge, SVR is just for regression tasks and not classification ones
- \Phi(x) in eq. 4 lacks a proper mathematical definition

The quality of English language is fine, apart from some minor issues.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Authors contribute to the Multi-output Regression Indoor Localization Algorithm based on Hybrid Grey Wolf Particle Swarm Optimization. The proposed localization framework is based on CSI and a novel fingerprint localization algorithm utilizing Hybrid Grey Wolf Particle Swarm Optimization. It's impressive.

1. Justification for selecting CSI over other distance measurement techniques such as RSSI, TDOA, etc., should be provided.

2.In your experiment, you used 50 reference nodes. Consider discussing the impact of the number of reference nodes on localization accuracy and include an analysis of this aspect.

3. Justification for selecting algorithms like FIFS, C-map, and LCAF for comparison with your proposed algorithms is needed. Also, addresses why deep learning-based algorithms were not included in the comparison.

3. Discuss how your algorithm might perform when applied to other wireless technologies like LoRa, BLE, Zigbee, and any considerations specific to these technologies.

4. Only one dataset was used in the experiment, which might not sufficiently demonstrate the generalization ability and robustness of the proposed methods. Recommend conducting experiments on multiple datasets to enhance the credibility of the results.

5.It's advisable to provide the final parameter values used in the machine learning algorithms in the conclusions.

Moderate editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

- Usually, Wi-Fi is used in the positioning in medium-sized environments such as a single- or multi-storey building where obstacles are ubiquitous. However, this study considers the positioning of a single target in a small area of 9x4m2 with light-of-sight condition. With these hypotheses, many other techniques such as camera, ultrasound, LIDAR, UWB may be more suitable. Therefore, the authors should explain what applications are aimed in this study.

- The PSOGWO serves as an optimization method that helps to search for the best estimated location that minimizes the loss function. This process is used only in the offline training phase, so the performance is not very crucial, and the reason to apply PSOGWO is weak except that it is original. In principle, any other global optimization method like GA, PSO,... can be used here when running them long enough, all of them would result in similar hyperparameters. But in Fig. 11, 3 methods converged to 3 different fitness values, then something must be not correct here.

- Otherwise, the method introduced in this paper is original and is a contribution that can be applied to other RF transceivers.

- Many math symbols are missing on page 11.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The updates addressed all my comments, and I don't have further comments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf