<p>In this work, we investigate device-to-device (D2D) mobile edge computing (MEC) with <i>intelligent reflecting surface (IRS)</i> aiding both D2D and device-to-MEC links to reduce energy and delay for user devices. We formulate a non-convex computation offloading problem that maximizes computing efficiency over latency under mixed integer, linear, and non-linear constraints and jointly optimize offloading decisions, partial offloading ratio, edge computing frequency, transmit power, and IRS phase shifts via a fully connected deep neural network (DNN) trained with a Lagrange-dual loss function that embeds the objective and inequality constraints. Extensive simulations across varying maximum task sizes (0.5-3 Mbits), edge computing capacities (10-90 Mcycles/s), and number of IRS elements (10-30) show that the proposed method consistently achieves solutions within 5-10% of near-global optimal solutions obtained by exhaustive search (ES) and outperforms gradient search (GS), fixed-offloading DNN and REINFORCE algorithm that achieves approximately 50% higher computing efficiency at <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$N_1=20$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>N</mi> <mn>1</mn> </msub> <mo>=</mo> <mn>20</mn> </mrow> </math></EquationSource> </InlineEquation> users. The proposed DNN exhibits 3-5% energy consumption gap from ES while maintaining lower computational complexity of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$O(12N_cN^2Ms)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>O</mi> <mo stretchy="false">(</mo> <mn>12</mn> <msub> <mi>N</mi> <mi>c</mi> </msub> <msup> <mi>N</mi> <mn>2</mn> </msup> <mi>M</mi> <mi>s</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> versus <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$O(\omega ^{-2})$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>O</mi> <mo stretchy="false">(</mo> <msup> <mi>ω</mi> <mrow> <mo>-</mo> <mn>2</mn> </mrow> </msup> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> for GS, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$O((2Q_s)^{(N+M)})$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>O</mi> <mo stretchy="false">(</mo> <msup> <mrow> <mo stretchy="false">(</mo> <mn>2</mn> <msub> <mi>Q</mi> <mi>s</mi> </msub> <mo stretchy="false">)</mo> </mrow> <mrow> <mo stretchy="false">(</mo> <mi>N</mi> <mo>+</mo> <mi>M</mi> <mo stretchy="false">)</mo> </mrow> </msup> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> for ES, and <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$O(\iota ^{-4})$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>O</mi> <mo stretchy="false">(</mo> <msup> <mi>ι</mi> <mrow> <mo>-</mo> <mn>4</mn> </mrow> </msup> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> for REINFORCE, which exhibits inference times on the order of milliseconds, compared to the computationally intractable costs associated with ES. These results show that the proposed algorithm for IRS-assisted partial D2D/MEC offloading is an effective and practical approach for jointly enhancing computing efficiency and latency in next-generation wireless systems.</p>

Computing efficiency and latency optimization in intelligent reflecting surface-enhanced device-to-device mobile edge computing via lagrange-dual deep learning

期刊论文
出版模式：

开放获取
发表日期：

2026年1月10日

查看全文

Kimchheang Chhea,
Jung-Ryun Lee

摘要

In this work, we investigate device-to-device (D2D) mobile edge computing (MEC) with intelligent reflecting surface (IRS) aiding both D2D and device-to-MEC links to reduce energy and delay for user devices. We formulate a non-convex computation offloading problem that maximizes computing efficiency over latency under mixed integer, linear, and non-linear constraints and jointly optimize offloading decisions, partial offloading ratio, edge computing frequency, transmit power, and IRS phase shifts via a fully connected deep neural network (DNN) trained with a Lagrange-dual loss function that embeds the objective and inequality constraints. Extensive simulations across varying maximum task sizes (0.5-3 Mbits), edge computing capacities (10-90 Mcycles/s), and number of IRS elements (10-30) show that the proposed method consistently achieves solutions within 5-10% of near-global optimal solutions obtained by exhaustive search (ES) and outperforms gradient search (GS), fixed-offloading DNN and REINFORCE algorithm that achieves approximately 50% higher computing efficiency at $N_1=20$ $N_{1} = 20$ users. The proposed DNN exhibits 3-5% energy consumption gap from ES while maintaining lower computational complexity of $O(12N_cN^2Ms)$ $O (12 N_{c} N^{2} M s)$ versus $O(\omega ^{-2})$ $O (ω^{- 2})$ for GS, $O((2Q_s)^{(N+M)})$ $O ({(2 Q_{s})}^{(N + M)})$ for ES, and $O(\iota ^{-4})$ $O (ι^{- 4})$ for REINFORCE, which exhibits inference times on the order of milliseconds, compared to the computationally intractable costs associated with ES. These results show that the proposed algorithm for IRS-assisted partial D2D/MEC offloading is an effective and practical approach for jointly enhancing computing efficiency and latency in next-generation wireless systems.