2026-10-07T11:00:23Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32826350 2026-10-01T16:08:39Z portal_1172 item_type_8 month_year_10_2026

Outage performance and calibration of ML-assisted resource allocation for next-generation wireless systems Rashika Raina (24304568) PUREID: 689577690 Blockage prediction custom loss function calibration greedy resource allocation machine learning novel loss function outage loss function outage prediction outage probability optimization resource allocation wireless communication wireless systems reliability URLLC reliability diagram artificial intelligence deep learning LSTM approximations sinc bessel covariance correlation decorrelation rayleigh fading rician fading optimal classifier reproducible AI trustworthy AI isotonic regression platt scaling post-processing calibration logarithmic binning linear binning Next-generation wireless networks, including 6G systems, require intelligent, adaptive strategies to mitigate link failures under dynamic channel conditions. While machine learning (ML) is a key enabler for such strategies, conventional ML approaches often yield limited performance gains and may fail to meet stringent reliability demands. This thesis develops novel ML solutions for intelligent resource allocation by predicting and avoiding link deterioration caused by random channel fluctuations. An ML-assisted resource allocation system is studied to anticipate reliability degradation in time-varying wireless environments, with outage probability (OP) serving as the metric for resource allocation error. As part of this framework, the predictor is trained using an outage loss function (OLF) specifically designed for this system. This work first analyses the outage performance of the system over Rayleigh fading channels, proposing closed-form analytical approximations and deriving conditions for optimal classifier performance. Results show that models trained via the OLF significantly outperform those trained with conventional binary cross-entropy. The thesis further investigates the calibration of the ML-based outage predictor, evaluating the reliability of its probability estimates. Histogram-based reliability diagrams with logarithmic binning are proposed to better characterise calibration in the low probability region relevant to high-reliability operation. Theoretical properties under perfect calibration are derived to guide threshold selection for meeting specific reliability requirements. Finally, the analysis is extended to Rician fading channels, which incorporate a dominant line-of-sight component representative of high-frequency wireless environments. The impact of the Rician K-factor on outage behaviour and calibration performance is examined, alongside tractable OP approximations. The thesis concludes by summarising the main contributions and outlining directions for future research.<br><br><i>Thesis is embargoed until 31 July 2027.</i> 2026-10-01T16:08:39Z Text Thesis 10.17034/32826350.v1 https://figshare.com/articles/thesis/Outage_performance_and_calibration_of_ML-assisted_resource_allocation_for_next-generation_wireless_systems/32826350 All Rights Reserved Open Access after 2027-07-31