A federated data-driven evolutionary algorithm for expensive multi-/many-objective optimization

Abstract

Data-driven optimization has found many successful applications in the real world and received increased attention in the field of evolutionary optimization. Most existing algorithms assume that the data used for optimization are always available on a central server for construction of surrogates. This assumption, however, may fail to hold when the data must be collected in a distributed way and are subject to privacy restrictions. This paper aims to propose a federated data-driven evolutionary multi-/many-objective optimization algorithm. To this end, we leverage federated learning for surrogate construction so that multiple clients collaboratively train a radial-basis-function-network as the global surrogate. Then a new federated acquisition function is proposed for the central server to approximate the objective values using the global surrogate and estimate the uncertainty level of the approximated objective values based on the local models. The performance of the proposed algorithm is verified on a series of multi-/many-objective benchmark problems by comparing it with two state-of-the-art surrogate-assisted multi-objective evolutionary algorithms.

Towards Fairness-Aware Multi-Objective Optimization

“…For example, different distributions of training data may result in surrogates of an MOP with different levels of fidelity [50], which may produce biases in the performance and result in biased search towards uninterested regions. In federated optimization [46] where the data subject to privacy can be collected only in a distributed way, performance fairness [9], collaboration fairness [9], and model fairness [9] will also become an important issue. In addition, optimization processes with/out considering fairness will result in different solutions [37].…”

“…Another difference is the training data distributed on the local clients and cannot be directly manipulated in an ensemble way. The fundamental difference between the federated optimization and federated learning is that federated optimization aims to assist the optimization process in finding the global optima or PF of the corresponding single optimization problem [45] or MOP [46], respectively, while federated learning is to train highly accurate global model [187], [188].…”

“…Especially, we can find out that the fairness issues may also exist in federated optimization. For example, the accuracy of the local models of federated optimization varies because a local model may not be updated when the new candidate solution is located in the infeasible region of optimization problem [45], [46]. As a result, inaccurate models may mislead the optimization process.…”

“…Model fairness shows that there is no discrimination of the model against specific groups of people. Meanwhile, the above notions of fairness also exist in federated optimization [45], [46], where data subject to privacy restrictions is collected in a distributed way for data-driven optimization. So far, few studies have ever investigated the fairness-aware federated optimization.…”

A Secure Federated Data-Driven Evolutionary Multi-objective Optimization Algorithm

“…For example, federated BO (FBO) [7] is proposed to tackle a global optimization task under a federated framework. To fully utilize the data of clients, Xu et al [8], [9] propose a federated acquisition function within a federated data-driven EA framework. Although these studies are inspired by FL, the following challenges arise as the targets of federated learning and FBO are different.…”