Parameter estimation of an ecological system by a neural network with residual minimization training

So far a considerable number of studies have been carried out on estimating parameters and completing imperfect parts of dynamical equations of an ecological system from a set of observation data. This paper presents a new method for such purposes by using a multi-layer neural network with residual minimization training, where an object function composed of a sum of squared residuals of supposed governing equations at collocation points and squared deviations of the solutions from observation data is minimized. y, we applied this method to analysis of “model observation data” produced by numerical integration of equations of a predator–prey system. As a result the parameters included in the model equations are recovered with considerably high accuracy even when artificial random noises are added to the original solutions. Then, we applied it successfully to real time-series data of protozoan populations. It is concluded that the new method can be effectively applied to analyses of time-series data in the field of the ecological science.