Using feedforward neural networks to represent ecosystem dynamics for bioeconomic analysis

Abstract

We applied feedforward neural networks to represent ecosystem dynamics that are vital to bioeconomic analysis, ecosystem-based management, or what-if analysis regarding the underlying natural resources. Neural networks are flexible, universal function approximators, recognized for their ability to recover complex nonlinear relationships. In this paper, we treated outputs from an end-to-end Atlantis model as synthetic data and used them as training data for the neural networks. After learning the seasonal dynamics of a multispecies system, we forecasted system states with different sets of specified harvest policies using the trained networks. We demonstrate that neural networks can capture key dynamics in part of the ecosystem efficiently, and give fast updates of states that are needed for optimization and decision-making. The trained networks are reduced and more flexible systems compared to the large-scale simulator model, which is more costly to run and does not have a format allowing human actions in the form of feedback policies, or harvest control rules, i.e. decisions depending on states as well as time.