BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260930T021029Z UID:3D35AD1F-01EA-46C8-8E65-EDA3193900B7 DTSTART;TZID=America/New_York:20261021T103000 DTEND;TZID=America/New_York:20261021T120000 DESCRIPTION:Abstract\n\nCell-edge rate is how next-generation networks get judged\, and almost none of the utilities we actually optimize target it. Percentile programs do. They are also non-convex\, non-smooth and strongly NP-hard. This talk follows that one problem across four papers\, and at e ach step there is less of the algorithm left for a person to design.\n\nIt begins with sum-least-q-th-percentile rate maximization\, which puts the fraction of users you care about directly into the objective and recovers max-min and sum-rate at its two ends. Two fractional transforms make it bl ock-concave\, giving cyclic algorithms that reach stationary points orders of magnitude faster than SCA or SQP\, and the same idea carries over to M U-MIMO beamforming through a somewhat unexpected equivalence with a percen tile-WMSE problem. Work in progress moves all of this to cell-free massive MIMO\, where coherent joint transmission breaks the concavity and a chang e of variables puts it back.\n\nThen the solver goes away. A neural networ k trained directly on the non-smooth utility\, differentiating through the sort itself\, matches the optimum where one is known and runs three to fo ur orders of magnitude faster.\n\nThen the designer goes away too! Given a n immutable evaluator and a page of instructions\, an AI agent ran eighty- one experiments over twenty-six hours and came back with a model within ha lf a percent of the best solver we have\, at roughly 600× lower inference cost\, and with a small theorem attached: the parameterization it found i s exactly max-min optimal at the minimum percentile\, whatever the weights turn out to be.\n\nI will close on what is left for us to do\, and the pa th forward for us as wireless researchers.\n\nSpeaker Biography\n\nAhmad K han (M’IEEE) received the Ph.D. and M.A.Sc. degrees from the Department of Electrical and Computer Engineering at the University of Toronto\, Cana da\, in 2023 and 2017\, respectively\, and the B.S. degree in Electrical E ngineering with First Class Honours from King Fahd University of Petroleum & Minerals (KFUPM)\, Dhahran\, Saudi Arabia\, in 2015. During his graduat e studies\, he was awarded the Edward S. Rogers Fellowship\, the Wallberg Memorial Fellowship\, the Ontario Graduate Scholarship\, and the Queen Eli zabeth II Graduate Scholarship in Science and Technology.\n\nIn 2021\, he joined Ericsson Ottawa as a Machine Learning Engineer\, and since 2023 has held a dual role there as Project Manager and Research Scientist in AI RA N R&D. In 2024\, he was a visiting researcher at Ericsson Research in Stoc kholm\, Sweden\, and is an alumnus of the company’s prestigious Early Ca reer Program. He also holds a patent in wireless resource management.\n\nH is research interests include wireless resource management — specificall y beamforming and scheduling — and applied machine learning for next-gen eration wireless networks. He has published in IEEE Transactions on Signal Processing\, including a two-part paper\, IEEE Transactions on Communicat ions\, IEEE Transactions on Wireless Communications\, IEEE Signal Processi ng Letters\, IEEE GLOBECOM\, and IEEE Access\, among others. His work span s collaborations with academic institutions including Yale University and he also supervises graduate students and postdoctoral researchers at the U niversity of Toronto through a funded research collaboration between Erics son and the university.\n\nRoom: 314\, Bldg: Walter Light Hall\, 19 Union St\, Kingston\, Ontario\, Canada\, K7L 3N9\, Virtual: https://events.vtool s.ieee.org/m/580863 LOCATION:Room: 314\, Bldg: Walter Light Hall\, 19 Union St\, Kingston\, Ont ario\, Canada\, K7L 3N9\, Virtual: https://events.vtools.ieee.org/m/580863 ORGANIZER:chan-f@rmc.ca SEQUENCE:40 SUMMARY:IEEE Talk: The Last Wireless Algorithms Designed by Humans URL;VALUE=URI:https://events.vtools.ieee.org/m/580863 X-ALT-DESC:Description: <br /><div style="mso-element: para-border-div\; bo rder: none\; border-bottom: solid #B4B4B4 1.0pt\; mso-border-bottom-alt: s olid #B4B4B4 .75pt\; padding: 0cm 0cm 6.0pt 0cm\;">\n<p class="MsoNormal" style="border: none\; mso-border-bottom-alt: solid #B4B4B4 .75pt\; padding : 0cm\; mso-padding-alt: 0cm 0cm 6.0pt 0cm\; margin: 14.0pt 0cm 6.25pt 0cm \;"><strong><span style="font-size: 12.0pt\; font-family: 'Cambria'\,serif \; mso-fareast-font-family: Cambria\; mso-bidi-font-family: Cambria\; colo r: #1f3864\; text-transform: uppercase\; letter-spacing: 1.2pt\;">Abstract </span></strong></p>\n</div>\n<p class="MsoNormal" style="margin-bottom: 7 .0pt\; text-align: justify\; line-height: 105%\;">Cell-edge rate is how ne xt-generation networks get judged\, and almost none of the utilities we ac tually optimize target it. Percentile programs do. They are also non-conve x\, non-smooth and strongly NP-hard. This talk follows that one problem ac ross four papers\, and at each step there is less of the algorithm left fo r a person to design.</p>\n<p class="MsoNormal" style="margin-bottom: 7.0p t\; text-align: justify\; line-height: 105%\;">It begins with <em>sum-leas t-q-th-percentile rate maximization</em>\, which puts the fraction of user s you care about directly into the objective and recovers max-min and sum- rate at its two ends. Two fractional transforms make it block-concave\, gi ving cyclic algorithms that reach stationary points orders of magnitude fa ster than SCA or SQP\, and the same idea carries over to MU-MIMO beamformi ng through a somewhat unexpected equivalence with a percentile-WMSE proble m. Work in progress moves all of this to cell-free massive MIMO\, where co herent joint transmission breaks the concavity and a change of variables p uts it back.</p>\n<p class="MsoNormal" style="margin-bottom: 7.0pt\; text- align: justify\; line-height: 105%\;"><em>Then the solver goes away.</em> A neural network trained directly on the non-smooth utility\, differentiat ing through the sort itself\, matches the optimum where one is known and r uns three to four orders of magnitude faster.</p>\n<p class="MsoNormal" st yle="margin-bottom: 7.0pt\; text-align: justify\; line-height: 105%\;"><em >Then the designer goes away too!</em> Given an immutable evaluator and a page of instructions\, an AI agent ran eighty-one experiments over twenty- six hours and came back with a model within half a percent of the best sol ver we have\, at roughly 600&times\; lower inference cost\, and with a sma ll theorem attached: the parameterization it found is exactly max-min opti mal at the minimum percentile\, whatever the weights turn out to be.</p>\n <p class="MsoNormal" style="margin-bottom: 1.5pt\; text-align: justify\; l ine-height: 105%\;">I will close on what is left for us to do\, and the pa th forward for us as wireless researchers.</p>\n<div style="mso-element: p ara-border-div\; border: none\; border-bottom: solid #B4B4B4 1.0pt\; mso-b order-bottom-alt: solid #B4B4B4 .75pt\; padding: 0cm 0cm 6.0pt 0cm\;">\n<p class="MsoNormal" style="border: none\; mso-border-bottom-alt: solid #B4B 4B4 .75pt\; padding: 0cm\; mso-padding-alt: 0cm 0cm 6.0pt 0cm\; margin: 14 .0pt 0cm 6.25pt 0cm\;"><strong><span style="font-size: 12.0pt\; font-famil y: 'Cambria'\,serif\; mso-fareast-font-family: Cambria\; mso-bidi-font-fam ily: Cambria\; color: #1f3864\; text-transform: uppercase\; letter-spacing : 1.2pt\;">Speaker Biography</span></strong></p>\n</div>\n<p class="MsoNor mal" style="margin-bottom: 7.0pt\; text-align: justify\; line-height: 105% \;"><strong>Ahmad Khan</strong> (M&rsquo\;IEEE) received the Ph.D. and M.A .Sc. degrees from the Department of Electrical and Computer Engineering at the University of Toronto\, Canada\, in 2023 and 2017\, respectively\, an d the B.S. degree in Electrical Engineering with First Class Honours from King Fahd University of Petroleum &amp\; Minerals (KFUPM)\, Dhahran\, Saud i Arabia\, in 2015. During his graduate studies\, he was awarded the Edwar d S. Rogers Fellowship\, the Wallberg Memorial Fellowship\, the Ontario Gr aduate Scholarship\, and the Queen Elizabeth II Graduate Scholarship in Sc ience and Technology.</p>\n<p class="MsoNormal" style="margin-bottom: 7.0p t\; text-align: justify\; line-height: 105%\;">In 2021\, he joined Ericsso n Ottawa as a Machine Learning Engineer\, and since 2023 has held a dual r ole there as Project Manager and Research Scientist in AI RAN R&amp\;D. In 2024\, he was a visiting researcher at Ericsson Research in Stockholm\, S weden\, and is an alumnus of the company&rsquo\;s prestigious Early Career Program. He also holds a patent in wireless resource management.</p>\n<p class="MsoNormal" style="margin-bottom: 7.0pt\; text-align: justify\; line -height: 105%\;">His research interests include wireless resource manageme nt &mdash\; specifically beamforming and scheduling &mdash\; and applied m achine learning for next-generation wireless networks. He has published in <em>IEEE Transactions on Signal Processing</em>\, including a two-part pa per\, <em>IEEE Transactions on Communications</em>\, <em>IEEE Transactions on Wireless Communications</em>\, <em>IEEE Signal Processing Letters</em> \, <em>IEEE GLOBECOM</em>\, and <em>IEEE Access</em>\, among others. His w ork spans collaborations with academic institutions including Yale Univers ity and he also supervises graduate students and postdoctoral researchers at the University of Toronto through a funded research collaboration betwe en Ericsson and the university.</p> END:VEVENT END:VCALENDAR