Our Work

Mazkur maqolada yangi 60ГРЛ markali marganes-borli quyma po'latning 35ХМЛ markali xrom-molibdenli po'latga nisbatan qo'llanish imkoniyatlari o'rganilgan. Tadqiqotda po'latlarning kimyoviy tarkibi va m…

Mazkur maqolada yangi 60ГРЛ markali marganes-borli quyma po'latning 35ХМЛ markali xrom-molibdenli po'latga nisbatan qo'llanish imkoniyatlari o'rganilgan. Tadqiqotda po'latlarning kimyoviy tarkibi va mexanik xossalari qiyosiy tahlil qilindi. Xususan, cho'zilishdagi mustahkamlik, qattiqlik va zarbiy qovushqoqlik ko'rsatkichlari aniqlanib, legirlovchi elementlar hamda termik ishlov berishning material xossalariga ta'siri baholandi. Shuningdek, molibden va bor elementlaridan foydalanishning texnik va iqtisodiy jihatlari ko'rib chiqildi.

Ushbu maqolada mobil aloqa uzilgan yoki signal darajasi pasaygan hududlarda uchuvchisiz uchish apparatlari (UAV) yordamida 5G tarmog'i qamrovini adaptiv tiklash hamda UAVning optimal koordinatalarini …

Ushbu maqolada mobil aloqa uzilgan yoki signal darajasi pasaygan hududlarda uchuvchisiz uchish apparatlari (UAV) yordamida 5G tarmog'i qamrovini adaptiv tiklash hamda UAVning optimal koordinatalarini aniqlash masalalari tadqiq etilgan. Taklif etilayotgan tizim mavjud bazaviy stansiyalarning aloqa sifatini baholash, optimal magistral aloqa tarmog'i (backhaul) manbasini tanlash va UAVning uch o'lchamli (3D) joylashuvini aniqlash algoritmlariga asoslanadi. Tadqiqotda UAV platformasida 2×2 MIMO magistral aloqa kanali hamda 360° gorizontal qamrovni shakllantiruvchi uchta 2×2 MIMO sektor antennalaridan foydalanish modeli ko'rib chiqilgan. Optimal joylashuvni aniqlashda RSRP, RSRQ, SINR, qamrov maydoni, tarmoq o'tkazuvchanligi va energiya sarfi ko'rsatkichlari kompleks hisobga olingan. Shuningdek, tarmoq parametrlari va ishlash sharoitlari o'zgarganda UAV koordinatalarini dinamik qayta optimallashtirish usuli ishlab chiqilgan.

Ushbu tadqiqotda XLM-R transformer modeli asosida olingan semantik embeddinglar yordamida Telegram matnlarini guruhlash imkoniyatlari o'rganildi. Agglomerative Clustering, Gaussian Mixture Model (GMM)…

Ushbu tadqiqotda XLM-R transformer modeli asosida olingan semantik embeddinglar yordamida Telegram matnlarini guruhlash imkoniyatlari o'rganildi. Agglomerative Clustering, Gaussian Mixture Model (GMM) va MiniBatch K-Means algoritmlari bir xil ma'lumotlar asosida eksperimental taqqoslandi. Klasterlarning tuzilishi, algoritmlarning ishlash xususiyatlari va katta hajmdagi embeddinglar bilan ishlash imkoniyatlari baholandi. Natijalar XLM-R embeddinglari matnlarni semantik yaqinligi bo'yicha samarali guruhlashini hamda algoritm tanlovi ma'lumotlar hajmi va tadqiqot maqsadiga bog'liqligini ko'rsatdi.

This paper addresses the development of an algorithm for selecting an efficient method for classifying objects in computer network traffic. The study analyzes rule-based packet classification, matchin…

This paper addresses the development of an algorithm for selecting an efficient method for classifying objects in computer network traffic. The study analyzes rule-based packet classification, matching types, including exact, prefix, and range matching, as well as the priorities assigned to these rules. The performance of various classification algorithms is evaluated in terms of computational time and memory consumption, and the algorithms are categorized into four main approaches. In the course of the study, Tuple Space Search, Hierarchical Tries, HyperCuts, and Aggregate Bit Vector algorithms are selected, and their suitability and efficiency for real-time systems are substantiated. Furthermore, the developed algorithm enables multi-criteria analysis of network packets, determination of the final class based on rule priorities, and detection of anomalies in network traffic.

A synthetic dataset generator that decouples cluster formation from class labels to evaluate cluster-label matching algorithms and measures.

Limitations of off-the-shelf software data collection platforms (e.g., Qualtrics, Survey Monkey) can possibly affect the types of survey data collected by higher education (HE) institutional data…

Limitations of off-the-shelf software data collection platforms (e.g., Qualtrics, Survey Monkey) can possibly affect the types of survey data collected by higher education (HE) institutional data professionals. Overcoming this limitation could be beneficial for increasing the quality of results obtained when striving to support HE stakeholders. Consideration of creative and novel approaches to collect data in a way that overcomes common platform limitations has historically been constrained by steep software learning curves. R is a language often used at HE institutions for working with data. Generative AI (GAI) conversational coding paired with the R Shiny framework allows for the rapid prototyping of proof-of-concept (POC) web applications by HE professionals with limited software development experience, going beyond the limitations of existing platforms and possibly better supporting the specific contextual needs of HE stakeholders. Demonstrated in this work is a conversational coding example using a design science research framework applied to a hypothetical HE recruitment scenario. A novel survey data collection web application artifact is produced as an illustrative applied example to understand in what ways a conversationally coded app can extend the capabilities of HE data professionals beyond features offered by existing data collection platforms. The result is a functional POC and perhaps offers a fresh perspective on what is possible by HE data professionals with limited programming experience.

This deposit contains a pre-registered study plan, analysis code, fixed input datasets, and design simulations for research on how post-training changes the geometry of language-model representations.…

This deposit contains a pre-registered study plan, analysis code, fixed input datasets, and design simulations for research on how post-training changes the geometry of language-model representations. It records the hypotheses, experimental settings, and analysis procedures before any main-study data were generated. Exploratory pilot analyses are disclosed in the plan. The deposited materials will remain non-public until their planned release. Version 1.1 (26 September 2026) adds Addendum A1 (addendum_A1.zip): analyses specified after the registered results were known and before these analyses were run. The registered plan, code and inputs of version 1.0 are unchanged and remain available under that version's DOI, 10.5281/zenodo.22858453. SHA-256 of addendum_A1.zip: 2edea85bebeaee809e09fc351233188a00145c02e4e91a023bef3a371c2c0e3b Version 1.2 (26 September 2026) adds Addendum A2 (addendum_A2.zip): an out-of-sample test, on the Tulu 3 / 3.1 8B and OLMo 2 7B post-training checkpoints, of the channel concentration suggested by Addendum A1. Specified after the results of Addendum A1 were known and before any hidden state of these families was extracted; before deposit, public weights were downloaded and the unchanged extractor was checked on three invented sentences, not on the study's passages. The registered plan and Addendum A1 are unchanged. SHA-256 of addendum_A2.zip: 294735750694aba7d9d9873d8e30f31e184a70b11e1112fc2dba755d886e8bf9Version 1.3 (26 September 2026) adds Addendum A3 (addendum_A3.zip): replication of the registered primary test and of the shape of the depth profile on OLMo 2 7B and on Tulu 3 / 3.1 8B (Llama-3.1-8B base). Specified after the registered results and both earlier addenda were known and before any response of these families was generated for the study's prompts; before deposit, public weights were downloaded and the unchanged generation and extraction code was checked on two invented prompts. The registered plan and Addenda A1 and A2 are unchanged. SHA-256 of addendum_A3.zip: ce827291c2dcac93fde58209f807687a3a99c9b0e999ef9d48b146b16886645cVersion 1.4 (27 September 2026) adds Addendum A4 (addendum_A4.zip): the registered supporting analysis SA3 (text versus weights, by generator-by-reader replay) on OLMo 2 7B and Tulu 3.1 8B, two families post-trained with the same recipe on different base models. Specified after the registered results and Addenda A1 to A3 were known and before any raw-format response of these families was generated; before deposit, the unchanged generation, extraction and replay code was checked on two invented prompts. The registered plan and Addenda A1 to A3 are unchanged. SHA-256 of addendum_A4.zip: 43179f7bfd590e5085a466ea8b8239401425204cb001f20f40432d9a15f1a282 Version 1.5 (1 October 2026) adds Addendum A5 (addendum_A5.zip). It tests whether the post-trained model’s lower sampling entropy explains the within-input fall. The test uses the registered primary contrast in OLMo 3 7B, Qwen2.5-7B, OLMo 2 7B and Tulu 3.1 8B. For each pair of models, the post-trained model is sampled at temperature 1. The base model is sampled at a temperature chosen so that its sampling entropy matches that of the post-trained model.The test was specified after the registered results, Addenda A1 to A4 and three exploratory checks were known. At that point, no base-model responses had been generated at a calibration temperature.Before deposit, the generation and extraction code was checked using the calibration settings on two invented prompts, without changing the code. The comparison code also reproduced the registered primary contrast.The registered plan and Addenda A1 to A4 are unchanged.SHA-256 of addendum_A5.zip: 1fc1c89febdac37185e73eabc3164e5c0260d52b8d45d519fa3ebba4989b4fc2Version 2.0 (October 2026) adds the plan of Study 2 (study2.zip): a confirmatory study on 300 fresh prompts in five model families (the four of Study 1, and Llama-3.1-8B -> Llama-3.1-8B-Instruct) that asks whether the narrowing of the answers to a prompt after post-training is more than what a narrower sampling distribution, matched in mean entropy, produces, and whether it lies in the answers or in how the model reads them. Registered before any Study 2 answer was generated: H1, the narrowing replicates at T = 1; H2a, at matched sampling entropy over three entropy levels (1, 1.25 and 1.5 times the post-trained model's entropy at T = 1), the post-trained model is narrower than its base model under the Tulu 3 recipe (OLMo 2, Tulu 3.1) and equivalent within 0.25 nats otherwise (OLMo 3, Qwen2.5, Llama-3.1 with Meta's recipe), with tests of difference and of equivalence in every family; H2b, the reader part of a writer-by-reader crossing at each level; H2c, the Tulu 3 recipe narrows more than Meta's on the same base model; H3, H2a against an on-task base model prompted with URIAL; H4, the across-input dimension does not fall by more than 0.25 nats (late depths without the last three blocks, raw and standardised states). Before deposit the prompts were drawn, the weights of Llama-3.1-8B-Instruct were downloaded, and the code was checked on invented prompts and on synthetic numbers. The plan of Study 1 and Addenda A1 to A5 are unchanged. SHA-256 of study2.zip: c8e20e60a93725354a8c49d1c65dd0d45e5260ce31cb0d67289684389b52eaafVersion 2.1 (3 October 2026) records two deviations from the Study 2 plan, both decided before any Study 2 result was opened (study2_v2.1.zip); the plan (v2.0) is unchanged. S2-D1: the run was split by family over two machines (spark-02: OLMo 3, Qwen2.5; spark-01: OLMo 2, then Tulu 3.1 and Llama-3.1). Each part applies the registered rules to its own families with the others entered at p = 1, so its corrected p-values are upper bounds; scripts/36_study2_merge.py joins the unadjusted rows and recomputes Holm's correction over the five families (ten tests for H4) with the registered functions of scripts/33_study2.py, estimating nothing, and a test checks that the result equals one run on all five families. S2-D2: the Study 2 configurations name no judge model, so the registered labelling step stops; scripts/37_study2_judge.py runs the same labelling with the Study 1 judge (Mistral-7B-Instruct-v0.3, commit c170c708), as the plan specifies, and the labels are reported without a rule. SHA-256 of study2_v2.1.zip: fc11646203cda7ea8bee638488ac58f682d0aea9dc4dcd487488891c61b3a3acVersion 2.2 (4 October 2026) records deviation S2-D3, decided before any Study 2 result was opened (study2_v2.2.zip);the plan (v2.0) is unchanged. At raised temperature Tulu 3.1 sampled token ids outside the vocabulary of its base model, Llama-3.1-8B (5 and 7 of 9,600 answers at entropy levels 2 and 3), and the writer-by-reader crossing stopped when the base model read those answers. scripts/34_study2_replay.py now re-tokenises such an answer from its text, the fallback it already used for special tokens; all other answers are read as before. The crossing of Tulu 3.1 at levels 2 and 3 and the rules of Tulu 3.1 and Llama-3.1 were run again (scripts/rerun_study2_s2d3.sh), keeping the crashed outputs. The record also corrects S2-D2: the OLMo 3 and Qwen2.5 configurations named the Study 1 judge, so only OLMo 2, Tulu 3.1 and Llama-3.1 needed it. SHA-256 of study2_v2.2.zip: 51c4be75bbac60c3b102bb1881f90f2ceddbc139be9b5d36b1d9fd6c582da455

Supplementary material of the paper "An Open-Source Data Generator for Joint Air Traffic Flow and Capacity Management Benchmarking" by A. Beiser, T. Gräupl, G. Trausmuth and S. Woltran, submitted…

Supplementary material of the paper "An Open-Source Data Generator for Joint Air Traffic Flow and Capacity Management Benchmarking" by A. Beiser, T. Gräupl, G. Trausmuth and S. Woltran, submitted to the Journal of Open Aviation Science (proceedings of the 14th OpenSky Symposium). preprint.pdf is the full paper with its two appendices, the differences from the first release of the datasets and the supplement to the statistical analysis. It is a preprint and has not yet been peer reviewed. statistical_analysis_code.zip holds the Python code and the result files behind the paper's statistical validation (Section 5, the hold-out paragraph of Section 6, Appendix 1 and Appendix 2). Its script reproduce_paper.sh recomputes the measurement runs, the classifier runs and the sampling-floor sweep from the published files alone; rerun that way, all results were identical to the shipped ones. The code's README.md maps every table and figure of the paper to the script and command that produce it. validation_instances_grid_regions.zip and validation_instances_xplane_regions.zip hold the instances of the validation that the two instance records do not contain: the additional seeds of the ten-seed scenarios and the hold-out sets generated from demand models fitted on the odd days of June 2019. Licences: the preprint is CC-BY-4.0, the code MIT, the validation instances of the grid regions CC-BY-4.0 and those of the regions built on X-Plane waypoints GPL-2.0-or-later (see LICENSING.md). The datasets: 10.5281/zenodo.23039645 (large-scaling), 10.5281/zenodo.23039835 (small-scaling); benchmark results: 10.5281/zenodo.23039865.

Results of 12 methods for joint Air Traffic Flow and Capacity Management (ATFCM) on all 6,920 instances of the V2 large-scaling and small-scaling benchmark datasets: one run per method and instance, 8…

Results of 12 methods for joint Air Traffic Flow and Capacity Management (ATFCM) on all 6,920 instances of the V2 large-scaling and small-scaling benchmark datasets: one run per method and instance, 83,040 runs, each limited to 1800 s and 35 GiB. The methods are nine configurations of the ASPaeroFlow heuristic (rerouting, delay and dynamic sectorization in different combinations, a CASA-like slot allocation and a sequential variant), a mixed-integer model solved with Gurobi, and two exact answer set programming encodings solved with clingo, all from the open-source ASPaeroFlow-Optimizer. runs.csv gives the outcome, runtime, memory and final objective values of every run (overload, arrival delay, sector number, sector changes, reroutes, reconfigurations), its validation status and sector metrics in the time windows of the formal model; tables.zip gives the same per problem and metric. The solution matrices of the 46,258 finished runs with a valid matrix are in five archives (18 GB), split by dataset and licence. Every solution matrix was checked independently against its instance: the six objectives were recomputed and compared with the values the method reported, and the hard constraints of the formal model were checked (graph edges and flight times, filed origins, destinations and departure times, aircraft rotations, sector partitions). No run disagrees with its reported objectives, and no complete matrix violates a constraint; validation.zip holds the check of every run. Licences: tables CC-BY-4.0; each solution matrix carries the licence of its instance (CC-BY-4.0, or GPL-2.0-or-later for the four regions built on X-Plane waypoints); scripts MIT (see LICENSING.md). README.md describes the methods, the setup, the metrics and the validation. Instances: 10.5281/zenodo.23039645 (large-scaling), 10.5281/zenodo.23039835 (small-scaling).

Version 2 of the small-scaling benchmark dataset for joint Air Traffic Flow and Capacity Management (ATFCM): 200 small, tightly capacitated instances, small enough to be solved to optimality by exact …

Version 2 of the small-scaling benchmark dataset for joint Air Traffic Flow and Capacity Management (ATFCM): 200 small, tightly capacitated instances, small enough to be solved to optimality by exact methods, so that heuristics can be compared against proven optima. The instances were generated from real June 2019 traffic of the OpenSky Network with the open-source data generator ASPaeroFlow-DataGenerator v2.0.0. The dataset covers 5 regions (EAST-ASIA-3x3, CENTRAL-EUROPE-5x5, INDIA-4x10, USA-7x7, MAJOR-EUROPE-10x10), all on rectangular grids of synthetic navpoints, with 10 demand levels from 10 to 100 flights, 4 seeds and hourly timesteps. Capacities are set so that on average 31.8 % of the sectors are over capacity. Every instance holds the navigation graph, the sectors with their capacities and allocation, the airports, the aircraft and their filed flights for one 24-hour day, and an instance_info.json with its parameters, generator commit and licence. The record also contains the unparsed output of the generator and its configurations. Compared with the first version, used in the LPNMR 2026 paper, the generator's demand model no longer discards the flights that cross UTC midnight, and destinations are drawn depending on origin and time of day. Licences: the instances are CC-BY-4.0, code is MIT. README.md describes the file formats, the formal model and how to cite it. The generator and the dataset are described in the paper "An Open-Source Data Generator for Joint Air Traffic Flow and Capacity Management Benchmarking" (preprint: 10.5281/zenodo.23039874); its large-scaling companion is 10.5281/zenodo.23039645 and benchmark results on both are 10.5281/zenodo.23039865.