Abstract
In federated learning (FL), to minimize the convergence time while reaching a desired accuracy, it is important to select appropriate mobile devices (MDs) as participants with the considerations of their characteristics (e.g., data distribution, channel condition, and computing power). In this article, we first formulate a Markov decision process-based MD selection problem in which the increased accuracy per unit-time is maximized. To solve the formulated problem without any prior knowledge of the environment, a deep Q-network (DQN) algorithm can be exploited. However, storing previous experiences into the replay memory for DQN consumes an increased time since the FL server needs to conduct a number of actual FL procedures and observe the resulted rewards. To address this problem, we proposed a virtual experience-based MD selection algorithm (VE-MSA). In VE-MSA, the FL server generates virtual experiences (especially reward) without any actual FL procedures by using two neural networks approximating the round time and the increased accuracy in a round according to the selected MDs, respectively. Evaluation results demonstrate that the derived optimal policy can achieve a target accuracy within the shortest time among comparison schemes.
Original language | English |
---|---|
Pages (from-to) | 1-10 |
Number of pages | 10 |
Journal | IEEE Systems Journal |
DOIs | |
Publication status | Accepted/In press - 2022 |
Keywords
- Client selection
- Computational modeling
- Convergence
- Data models
- deep Q-network (DQN)
- experience relay
- federated learning (FL)
- Neural networks
- Servers
- straggler problem
- Training
- Training data
ASJC Scopus subject areas
- Control and Systems Engineering
- Information Systems
- Computer Science Applications
- Computer Networks and Communications
- Electrical and Electronic Engineering