Lokálně spojené echo state sítě pro predikci časových řad

Matzner, Filip

Echo State Networks (ESNs) are a class of recurrent neural networks in which only a small readout regression layer is trained, while the weights of the recurrent network, termed the reservoir, are randomly assigned and remain fixed. Although ESNs have shown potential in time series forecasting (TSF), they often fail to match state-of-the- art deep learning approaches on real-world benchmarks. In this thesis, we introduce Locally Connected ESN (LCESN), a novel ESN variant improving scalability, stability, and accuracy in TSF tasks. Its local topology significantly reduces asymptotic time and space complexities compared to the conventional ESN, enabling substantially larger networks. The forced memory mechanism enhances their ability to capture long-range dependencies without destabilizing network dynamics. Another key contribution is a novel hyperparameter tuning method, critical for the sensitive dynamic system inherent to ESNs. This thesis is structured as a commented collection of three original research papers, each providing new insights into recurrent neural networks for TSF. Our re- sults demonstrate that well-tuned LCESNs outperform conventional ESNs and achieve competitive accuracy with state-of-the-art models. These findings reaffirm the relevance of recurrent architectures in modern TSF research,...

guest :: login Digital Repository
		Search		Submit		Help		About