Recognition of human activities based on decision optimization model

A fast adaptive beamformer with sidelobe control based on gradient descent ascent

“…When M = 1, our results recover the convergence results of Smoothed-AGDA in the centralized setting [Yang et al, 2022b]. Similar to Yang et al [2022b], we can also translate an (ϵ, ϵ/ √ κ)-stationary point of f to an ϵ-stationary point of Φ under the federated setting, as stated below.…”

“…smoothing technique, and prove that it has a faster convergence rate for centralized nonconvex-concave problems compared with GDA. Yang et al [2022b] then prove that Smoothed-AGDA and its stochastic version Stochastic Smoothed-AGDA also have faster convergence rates for centralized nonconvex-PL (Polyak-Lojasiewicz) problems compared with GDA (SGDA). So a natural question arises: Can we utilize smoothing techniques and design a faster algorithm for federated nonconvex minimax optimization?…”

“…We state some common assumptions that will be used throughout the paper. They are commonly used in (federated) minimax optimization; e.g., [Yang et al, 2022b, Deng and Mahdavi, 2021.…”

“…Inspired by the success of Smoothed-AGDA in the centralized setting , Yang et al, 2022b, we propose FESS-GDA, which is compactly presented in Algorithm 1, for the federated minimax optimization problem. We consider a system with M clients and one central server.…”

“…We consider a setting similar to Yang et al [2022b], Loizou et al [2020, using a Wasserstein GAN [Arjovsky et al, 2017] to approximate a one-dimensional Gaussian distribution in the federated setting. We first randomly generate a synthetic dataset of n = 10000 datapoints z sampled from a normal distribution with zero mean and unit variance and their corresponding real data x real = μ + σz, where μ = 0, σ = 0.1.…”