SAIFR:2135

Time-series forecasting using recurrent neural networks and Takens’ Theorem

(2020). Time-series forecasting using recurrent neural networks and Takens’ Theorem. ICTP South American Institute for Fundamental Research. https://scivideos.org/index.php/ictp-saifr/2135

Time-series forecasting using recurrent neural networks and Takens’ Theorem. ICTP South American Institute for Fundamental Research, Mar. 02, 2020, https://scivideos.org/index.php/ictp-saifr/2135 

          @misc{ scivideos_SAIFR:2135,
            doi = {},
            url = {https://scivideos.org/index.php/ictp-saifr/2135},
            author = {},
            keywords = {ICTP-SAIFR, IFT, UNESP},
            language = {en},
            title = {Time-series forecasting using recurrent neural networks and Takens{\textquoteright} Theorem},
            publisher = { ICTP South American Institute for Fundamental Research},
            year = {2020},
            month = {mar},
            note = {SAIFR:2135 see, \url{https://scivideos.org/index.php/ictp-saifr/2135}}
          }
Laís Alves
Talk numberSAIFR:2135
Source RepositoryICTP – SAIFR
Collection
Talk Type Conference
Subject

Abstract

Artificial Neural Networks (ANNs) have been demonstrated to be an excellent method for dealing with various kind of tasks, such as image classification and natural language processing. These ANNs can also be used for regression since they are considered to be an universal function approximator, thanks to it's great capacity of dealing with non linear tasks. On the other hand, Floris Takens have shown to be possible to access and reconstruct the underlying dynamics of a system in the space state, starting from a single measured time series. In other words, Takens' theorem says it is possible to get information of a higher dimensional system from vectors build from a one-dimensional time series. This enables better time-series representations that can be fed into the ANNs. In this work, I will apply Echo State Networks (ESNs), one among the types of ANNs, to make forecasts of the dynamics of physical systems, such as Lorenz, Fokker-Planck and Vlasov Equations. I also pretend to apply these ESNs, together with Takens Vectors, into real time series, such as epidemics spread and financial series.