{"id":4818,"date":"2025-11-25T07:21:23","date_gmt":"2025-11-25T07:21:23","guid":{"rendered":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?page_id=4818"},"modified":"2025-11-25T07:28:13","modified_gmt":"2025-11-25T07:28:13","slug":"stacked-meta-learner-blueprint-for-rf-modulation-ensembles","status":"publish","type":"page","link":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?page_id=4818","title":{"rendered":"Stacked Meta-Learner Blueprint for RF Modulation Ensembles"},"content":{"rendered":"\n<div data-wp-interactive=\"core\/file\" class=\"wp-block-file\"><object data-wp-bind--hidden=\"!state.hasPdfPreview\" hidden class=\"wp-block-file__embed\" data=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/Stacked-Meta-Learner-Blueprint-for-RF-Modulation-Ensembles-Rev2.pdf\" type=\"application\/pdf\" style=\"width:100%;height:600px\" aria-label=\"Embed of Stacked Meta-Learner Blueprint for RF Modulation Ensembles Rev2.\"><\/object><a id=\"wp-block-file--media-1b0b9f50-fe8a-4758-9bae-3fb3e3275522\" href=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/Stacked-Meta-Learner-Blueprint-for-RF-Modulation-Ensembles-Rev2.pdf\">Stacked Meta-Learner Blueprint for RF Modulation Ensembles Rev2<\/a><a href=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/Stacked-Meta-Learner-Blueprint-for-RF-Modulation-Ensembles-Rev2.pdf\" class=\"wp-block-file__button wp-element-button\" download aria-describedby=\"wp-block-file--media-1b0b9f50-fe8a-4758-9bae-3fb3e3275522\">Download<\/a><\/div>\n\n\n\n<p>a concrete blueprint for enabling stacked<br>ensembles in the existing production path. We construct metafeatures from per-model logits and probabilities emitted by EnsembleMLClassifier, train logistic-regression and gradient-boostedtree meta-learners using cross-validated out-of-fold predictions,<br>and compare their behaviour to the current weighted voting<br>baseline. Our experiments on synthetic RF scenarios show that<br>properly cross-validated stacking yields up to 1.3 and 2.0 absolute<br>accuracy points over weighted voting for logistic and GBM metalearners respectively, while naive (non-CV) stacking overfits by<br>as much as 6.7 percentage points. We release a harness and<br>figure-generation scripts so the stacked path can be turned on or<br>off by configuration without modifying the LATEX.<br>Index Terms\u2014Automatic modulation classification, ensembles,<br>stacked generalization, meta-learning, RF machine learning.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Repository:<\/h2>\n\n\n\n<p><a href=\"https:\/\/github.com\/bgilbert1984\/Stacked-Meta-Learner-Blueprint-for-RF-Modulation-Ensembles\">bgilbert1984\/Stacked-Meta-Learner-Blueprint-for-RF-Modulation-Ensembles: Ensembles of heterogeneous architectures (SpectralCNN, SignalLSTM, ResNetRF, SignalTransformer) trade off accuracy, latency, and energy under realistic RF workloads. Majority and confidence- weighted voting are simple to implement and integrate cleanly with hierarchical routing and open-set policies.<\/a><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><a href=\"https:\/\/mastodon.social\/@Bgilbert1984\"><img data-opt-id=1549558125  fetchpriority=\"high\" decoding=\"async\" width=\"1024\" height=\"583\" src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:1024\/h:583\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-53.png\" alt=\"\" class=\"wp-image-4820\" srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:1024\/h:583\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-53.png 1024w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:300\/h:171\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-53.png 300w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:768\/h:437\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-53.png 768w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:1054\/h:600\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-53.png 1054w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/a><\/figure>\n\n\n\n<p>To include a Pareto plot in your paper &#8220;Ensemble Size vs Latency and Energy on CPU\/GPU for RF Modulation Ensembles,&#8221; I recommend adding it as Figure 3 in Section V (Results) or VI (Discussion). This visualization would explicitly show the trade-offs (e.g., accuracy vs. latency\/energy, with points colored by ensemble size k), highlighting &#8220;knees&#8221; and Pareto frontiers for CPU\/GPU deployments. Based on the paper&#8217;s data (e.g., accuracy plateauing at ~0.89 beyond k=4, latency growing linearly to p99=18.3ms at k=4), a sample plot could look like this simulated example derived from your empirical curves:<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img data-opt-id=1962656412  fetchpriority=\"high\" decoding=\"async\" width=\"685\" height=\"548\" src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:auto\/h:auto\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-54.png\" alt=\"\" class=\"wp-image-4826\" srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:685\/h:548\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-54.png 685w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:300\/h:240\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/11\/image-54.png 300w\" sizes=\"(max-width: 685px) 100vw, 685px\" \/><\/figure>\n\n\n\n<p>This plot (adapted from similar RF\/ML studies) uses scatter points for each k, a dashed frontier line connecting non-dominated points, and color-coding for size. It would strengthen your &#8220;budget plot&#8221; contribution by making deployment decisions more intuitive\u2014e.g., GPU&#8217;s flatter curve allows larger k at similar costs. If generating from your exact data, use matplotlib as in the code example I tested (outputs a PNG for inclusion in LaTeX via \\includegraphics). This addition would elevate the paper&#8217;s visual impact without exceeding page limits.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Pareto plot for ensemble size vs latency energy accuracy in RF modulation classification10<\/h3>\n\n\n\n<p><a href=\"https:\/\/www.researchgate.net\/figure\/Pareto-frontier-chart-classification-accuracy-vs-energy-usage-per-test-sample-for-models_fig8_394817535\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/www.researchgate.net\/figure\/Pareto-frontier-chart-classification-accuracy-vs-energy-usage-per-test-sample-for-models_fig8_394817535\" target=\"_blank\" rel=\"noreferrer noopener\">Pareto frontier chart classification accuracy vs. energy usage per test&#8230;<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/www.researchgate.net\/figure\/Pareto-frontier-chart-classification-accuracy-vs-energy-usage-per-test-sample-for-models_fig8_394817535\" target=\"_blank\" rel=\"noreferrer noopener\">Pareto frontier chart classification accuracy vs. energy usage per test sample for models in ablation study.<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/www.researchgate.net\/figure\/Pareto-frontier-chart-classification-accuracy-vs-energy-usage-per-test-sample-for-models_fig8_394817535\" target=\"_blank\" rel=\"noreferrer noopener\">researchgate.net<\/a><a href=\"https:\/\/www.mdpi.com\/2076-3417\/13\/14\/8511\" target=\"_blank\" rel=\"noreferrer noopener\">Predictive Modeling of Delay in an LTE Network by Optimizing the &#8230;This paper aims to investigate the training, testing and selection of the best predictive delay model for a VoIP service in a Long Term Evolution (LTE) networkmdpi.com<\/a><a href=\"https:\/\/codeocean.com\/explore\/e92087f6-1ce0-40a5-b08d-3f76aeadbefd?query=IEEE%252520Latin%252520America%252520Transactions&amp;page=1&amp;filter=all&amp;refine=journal\" target=\"_blank\" rel=\"noreferrer noopener\">IEEE Latin American ID6390-LCOE-Pareto_Front &#8211; Code OceanThe Pareto front is obtained for each of the variables that are defined as relevant in sensitivity. The optimization problem model is applied to the proposed &#8230;codeocean.com<\/a><a href=\"https:\/\/kastner.ucsd.edu\/wp-content\/uploads\/2020\/05\/admin\/raw2020-real_time_modulation_classification.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">[PDF] Real-time Automatic Modulation Classification using RFSoCOur work explores the practicalities of realizing a high throughput and low latency system implementation for doing AMC. It examines the trade- offs in a number &#8230;kastner.ucsd.edu<\/a><a href=\"https:\/\/pmc.ncbi.nlm.nih.gov\/articles\/PMC12216454\/\" target=\"_blank\" rel=\"noreferrer noopener\">Ensemble learning for biomedical signal classification &#8211; NIHBy achieving a classification accuracy of 95.4%, the ensemble framework outperformed traditional classifiers in capturing subtle diagnostic &#8230;pmc.ncbi.nlm.nih.gov<\/a><a href=\"https:\/\/ieeexplore.ieee.org\/iel8\/7361\/4427201\/11119040.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Enhancing the Accuracy of Visible Light Positioning Systems Using &#8230;This study aims to enhance VLP localization accuracy by processing high-dimensional input features using a Kolmogorov\u2013Arnold Network (KAN) and comparing its &#8230;ieeexplore.ieee.org<\/a><a href=\"https:\/\/arxiv.org\/html\/2501.06726v2\" target=\"_blank\" rel=\"noreferrer noopener\">Integrated Sensing and Edge AI: Realizing Intelligent Perception in 6G&#8230; accuracy, latency, and energy efficiency. It aligns with not only task-oriented communications of 6G but also the design principle of E2E perception &#8230;arxiv.org<\/a><a href=\"https:\/\/www.sciencedirect.com\/science\/article\/pii\/S1361841521000645\" target=\"_blank\" rel=\"noreferrer noopener\">Ultrasound image reconstruction from plane wave radio-frequency &#8230;A novel self-supervised US reconstruction method based on deep neural network. An encoder-decoder architecture produces high-quality US images.sciencedirect.com<\/a><a href=\"https:\/\/www.researchgate.net\/figure\/Pareto-optimization-among-90-NN-models-a-accuracy-vs-latency-b-accuracy-vs_fig12_393176241\" target=\"_blank\" rel=\"noreferrer noopener\">Pareto optimization among 90 NN models: (a) accuracy vs. latency &#8230;Download scientific diagram | Pareto optimization among 90 NN models: (a) accuracy vs. latency, (b) accuracy vs. memory, (c) memory vs. latency. from &#8230;researchgate.net<\/a><a href=\"https:\/\/link.springer.com\/article\/10.1007\/s10462-025-11317-4\" target=\"_blank\" rel=\"noreferrer noopener\">A comprehensive systematic literature review on artificial &#8230;Analysing nine performance metrics, Convolutional Neural Network was the most popular choice (20.6%) with an accuracy of 99% and SNR from 6-20dB &#8230;link.springer.com<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Pareto plot for ensemble size vs accuracy latency energy in RF modulation classification1<\/h3>\n\n\n\n<p><a href=\"https:\/\/link.springer.com\/article\/10.1007\/s10462-025-11317-4\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/link.springer.com\/article\/10.1007\/s10462-025-11317-4\" target=\"_blank\" rel=\"noreferrer noopener\">A comprehensive systematic literature review on artificial &#8230;<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/link.springer.com\/article\/10.1007\/s10462-025-11317-4\">https:\/\/link.springer.com\/article\/10.1007\/s10462-025-11317-4<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Executed python code19 lines<\/h3>\n\n\n\n<p>Python<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import matplotlib.pyplot as plt\nimport numpy as np\n\n<em># Sample data from paper critiques<\/em>\nk = &#91;1, 2, 4, 6, 8]\naccuracy = &#91;0.85, 0.87, 0.89, 0.891, 0.892]\nlatency_cpu = &#91;5, 10, 18, 25, 32]  <em># ms<\/em>\nenergy_cpu = &#91;10, 18, 30, 40, 50]  <em># mJ<\/em>\n\n<em># Plot Pareto: accuracy vs latency, colored by k<\/em>\nfig, ax = plt.subplots()\nscatter = ax.scatter(latency_cpu, accuracy, c=k, cmap='viridis')\nax.plot(latency_cpu, accuracy, 'k--', alpha=0.5)  <em># frontier line<\/em>\nax.set_xlabel('Latency (ms, CPU)')\nax.set_ylabel('Accuracy')\nax.set_title('Pareto Frontier: Accuracy vs Latency by Ensemble Size')\nplt.colorbar(scatter, label='Ensemble Size (k)')\nplt.savefig('pareto_plot.png')\nprint('Plot saved as pareto_plot.png')<\/code><\/pre>\n\n\n\n<p><a href=\"https:\/\/grok.com\/share\/bGVnYWN5_594f1534-142a-42f7-a260-71dc8c970dd3\">https:\/\/grok.com\/share\/bGVnYWN5_594f1534-142a-42f7-a260-71dc8c970dd3<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>a concrete blueprint for enabling stackedensembles in the existing production path. We construct metafeatures from per-model logits and probabilities emitted by EnsembleMLClassifier, train logistic-regression and gradient-boostedtree meta-learners using cross-validated out-of-fold predictions,and compare their behaviour to the current weighted votingbaseline. Our experiments on synthetic RF scenarios show thatproperly cross-validated stacking yields up to 1.3 and 2.0&hellip;&nbsp;<a href=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?page_id=4818\" rel=\"bookmark\"><span class=\"screen-reader-text\">Stacked Meta-Learner Blueprint for RF Modulation Ensembles<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":4820,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"googlesitekit_rrm_CAowgMPcCw:productID":"","neve_meta_sidebar":"","neve_meta_container":"","neve_meta_enable_content_width":"","neve_meta_content_width":0,"neve_meta_title_alignment":"","neve_meta_author_avatar":"","neve_post_elements_order":"","neve_meta_disable_header":"","neve_meta_disable_footer":"","neve_meta_disable_title":"","footnotes":""},"class_list":["post-4818","page","type-page","status-publish","has-post-thumbnail","hentry"],"_links":{"self":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/pages\/4818","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4818"}],"version-history":[{"count":3,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/pages\/4818\/revisions"}],"predecessor-version":[{"id":4827,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/pages\/4818\/revisions\/4827"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/media\/4820"}],"wp:attachment":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4818"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}