# Matching-Theory-Based Low-Latency Scheme for Multitask Federated Learning in MEC Networks 在現有MEC架構下,edge node可以有足夠資源來進行FL的aggregate同時其也足夠接近IoT device。藉由可以減少communication時間。 此篇主要貢獻: 1. 使用many-to-one( hospitals–residents) matchong problem來formulate 多個client與多個server間的配對問題 * client只能assign給一個server,但一個server可以assign給多個client 2. 使用incomplete preference list(IPL)來描述matching中每個player的preference * 在IoT環境下,不可能每個client與server都能夠完整了解全部人的狀況並且做出preference list for全部對象,因此更為實際的必會是incomplete * 假設preference會是strict order 3. 追求low-latency * 提出的algorithm目標是minimize整體multi-task FL的訓練時間(objective) MODEL(TIME): 1. total time = computation time + communication time 2. computation time  * 紅色: 表示在要達到**目標accuracy下所需要local update次數** (引用paper: *Federated learning over wireless networks: Optimization model design and analysis*) * 每round時間則是 **每輪update所需資料量*不同model隊單位資料處理量/node計算能力** (引用paper: *Self organizing federated learning over wireless networks: A socially aware clustering approach*) 3. communication time  transmission rate:  PROBLEM: * <font color="#f00">這裡我覺得不是很合理有點怪怪的,這讓我感覺只是一個特別掰出來的object function用來衡量是不是他所提出matching方式有做到很好。但我覺得應該事先提出optimized problem再來決定方式設計好不好</font> * <font color="#f00">此處的latency是全部相加,但實際狀況是只需要考量最慢的那個人就好</font> * <font color="#f00">此處考量的是所有model最終一起訓練時間,但實際應該該在乎的是在一個model中,是否有一個stragger且可能有些model本來就練比較久,應針對對個別model來reduce latency</font>   * r θi is introduced, which is denoted as the willingness for participation of the ith device. Intuitively, θi is determined by the remained energy percentage, i.e., θi = 1 when δ ≤ ci and vice versa. 1. **每輪global round執行一次matching** 2. **每輪global round都optimized,則整體latency也optimized** 3. **只要local update次數足夠就可以保證最終global有足夠accuracy** (引用paper: 1. Communication-efficient federated deep learning with layerwise asynchronous model update and temporally weighted aggregation 2. Federated learning based mobile edge computing for augmented reality applications) MATCHING: 1. IPL in the HR game 2. 作者提到,雖然演算法**允許每輪都重新配對,但這樣的配對會造成需要重新更新local端的model以及相關設定耗時間**(hyperparameter及loss function等等),因此我們要去找一個"stable matching"確保不會一值變動 1. stable in IPL * 簡單講不存在一組pair能夠block matching。而block表示裡面的player能夠找到更好的pair於現有情況中  ALGORITHM: 1. preference如何設定: 1. **client對server**: * 依據每個client與server的communication time(越小越喜歡) * client的energy resource必須大於一個值for server model(因為對client來說運算需要耗能,而行動裝置很在意電量,這對client是一個重要的考慮因素)  * com - > communication * N_i -> IoT device 2. **server對client**: * 依據每個client對server model的computation time(越小越喜歡)  * cmp -> computation * j -> edge node(server) 2. 演匴法流程(沒有看)  其他: <font color="#f00">此篇看起來實際沒有解決non-iid問題,在分配時所使用到preference list是依據時間等因素來排而非data distribution</font>
×
Sign in
Email
Password
Forgot password
or
By clicking below, you agree to our
terms of service
.
Sign in via Facebook
Sign in via Twitter
Sign in via GitHub
Sign in via Dropbox
Sign in with Wallet
Wallet (
)
Connect another wallet
New to HackMD?
Sign up