Abstract

Artificial intelligence has revolutionized various societal and business processes, offering faster and more cost-effective problem-solving capabilities. However, traditional machine learning approaches still have limitations, including their reliance on processing power and the challenges associated with centralized algorithms. These limitations result in higher user costs and negative environmental impacts. To overcome these challenges, decentralized and distributed environments, such as the cloud or the edge, can be utilized. This paper proposes a general method for addressing multi-objective optimization problems in adaptive and federated machine learning systems. We first analyze the inherent noise in adaptive and decentralized systems through extensive benchmarking to mitigate errors and ensure reliability. Next, we define multiple optimization criteria, including training time, resource utilization, and rewards related to available resources. We present a multi-objective optimization model specifically designed for improved federated machine learning on edge and cloud infrastructures. Lastly, we introduce an automated configuration approach for federated learning platforms using hyperparameter optimization. The evaluation shows that our method considerably improves training time and resource wastage with optimized rewards for adaptive federated learning systems.