分布式数据驱动优化场景下的云边协同进化优化算法数据集
收藏资源简介：
代理模型辅助的进化算法(surrogate-assisted evolutionary algorithms, SAEAs)近年来被提出用于解决数据驱动优化问题。多数现存的代理模型辅助的进化算法是为集中式优化而设计的，并未考虑物联网时代下数据分布于网络的边缘带来的挑战。为此，本研究提出云边协同进化优化算法（edge-cloud co-evolutionary algorithms，ECCoEAs）来解决分布式数据驱动优化问题（distributed data-driven optimization problems, DDOPs）。具体来说，本研究首先提出了一个云边协同进化优化算法的分布式框架。这个框架由一个通信机制、边缘模型管理和云模型管理三部分组成。这个通讯机制控制模型信息、终止信息、有效候选解和最终解的通讯顺序来防止边缘服务器和云服务器协作过程中死锁的产生。在边缘模型管理中，边缘代理模型是基于局部数据训练而成。这些局部数据包括了局部历史数据以及由协同进化生成的新解与其真实评估组成的数据。在云模型管理中，从边缘服务器接收到的边缘代理模型的黑盒预测函数被用于构造一个全局模型。这个模型进而辅助云服务器上的进化优化来寻找有效候选解。有效候选解一方面可能是问题最终的最优解，另一方面可以通过通讯机制传输给边缘服务器来指导边缘模型管理。此外，为了验证框架的通用性，本研究实现了两个ECCoEAs，分别是边缘云协同代理辅助差分进化算法（ECCo-SDE）和边缘云协同代理辅助分层粒子群优化算法（ECCo-SHPSO）。 本数据集首先介绍了所用的基准测试函数所拥有的独立同分布以及非独立同分布的两种数据分布情况下各个边缘节点的初始历史数据。本数据集然后介绍提出的两个ECCoEAs与其集中式版本对比在基准测试函数的两种数据分布情况进行的大量实验研究得到的原始数据进一步分析后得到的分析数据，以表明本研究提出的ECCoEAs能够有效地解决分布式数据驱动优化问题,达到与其集中式版本解决数据集中时的数据驱动问题时相似的性能，且ECCoEAs的分布式框架具有通用性。最后，本数据集介绍提出的ECCo-SHPSO在三个分布式聚类问题上进行大量实验研究得到的原始数据以及进一步分析后得到的分析数据，以表明提出的ECCo-SHPSO与两个分布式聚类算法对比有更高的求解性能。
Surrogate-assisted evolutionary algorithms (SAEAs) have been proposed in recent years to solve data-driven optimization problems. Most existing SAEAs are designed for centralized optimization, without considering the challenges brought by the distributed deployment of data at network edges in the Internet of Things (IoT) era. To address this issue, this study proposes edge-cloud co-evolutionary algorithms (ECCoEAs) to solve distributed data-driven optimization problems (DDOPs). Specifically, this study first proposes a distributed framework for ECCoEAs. The framework consists of three components: a communication mechanism, edge model management, and cloud model management. The communication mechanism controls the transmission order of model information, termination signals, valid candidate solutions, and final solutions to avoid deadlocks during the collaboration between edge servers and cloud servers. For edge model management, edge surrogate models are trained based on local data, which includes local historical data as well as data composed of newly generated solutions via co-evolution and their real evaluations. For cloud model management, the black-box prediction functions of edge surrogate models received from edge servers are used to construct a global model, which further assists evolutionary optimization on cloud servers to find valid candidate solutions. On one hand, these valid candidate solutions may be the final optimal solutions of the problem; on the other hand, they can be transmitted to edge servers via the communication mechanism to guide edge model management. Additionally, to verify the generality of the proposed framework, this study implements two instances of ECCoEAs: edge-cloud co-surrogate-assisted differential evolution (ECCo-SDE) and edge-cloud co-surrogate-assisted hierarchical particle swarm optimization (ECCo-SHPSO). This dataset first presents the initial historical data of each edge node under two data distribution scenarios, i.e., independent and identically distributed (IID) and non-independent and identically distributed (non-IID), for the adopted benchmark functions. Next, this dataset provides the analyzed data derived from further processing of the raw experimental results obtained from extensive comparative experiments between the two proposed ECCoEAs and their centralized counterparts, conducted under both IID and non-IID data distribution scenarios for the benchmark functions. These data demonstrate that the proposed ECCoEAs can effectively solve DDOPs, achieving performance comparable to that of their centralized versions when solving data-driven problems with centralized data, and that the distributed framework of ECCoEAs has good generality. Finally, this dataset presents the raw experimental data and further analyzed data obtained from extensive experiments of the proposed ECCo-SHPSO on three distributed clustering problems, which demonstrate that the proposed ECCo-SHPSO achieves higher solving performance when compared against two baseline distributed clustering algorithms.
