Time series-based bifurcation diagram reconstruction (Q1809178)

The main idea that the authors use is the ability of neural networks to learn general solutions to problems if given an appropriate set of specific examples. In the centre of interest is the problem of reconstructing bifurcation diagrams of dynamical systems without the knowledge of its functional form and its dependence of the parameter. By using time series at different parameter values a three-layer fully connected neural network is employed in the approximation of the map. The task of this network is to learn the dynamic of the system as function of the parameters from the time series. For a special class of maps (linear in parameter, LIP) a scheme in locating a linear subspace in the network's weight space with a qualitatively similar bifurcation structure as that of the map is discussed. Hereby the principal component analysis (PCA) plays an important role. With the predictor functions obtained by the neural network an algorithm in reconstructing the bifurcation diagrams for this class of maps is proposed by the authors. Other classes of predictors could also be used due to the flexibility of the algorithm.