Exploring local effects of economic activities using agent-based ecosystem models

We present an ecosystem simulator with three-dimensional models of topography and terrain. The simulator includes animal models that are individually controlled by deep reinforcement learning. The simulations take place in a game engine environment, which enables continuous observation of the ecosystem from any camera angle. The topographic models are generated from geographic data with altitudes and land cover type. The animal models combine three-dimensional conformation models and animation schemes with decision-making mechanisms trained with deep reinforcement learning in increasingly complex environments (curriculum learning). We show how this framework can be used for simulating the development of specific ecosystems with or without different forms of human intervention. In particular, we show how it can be used for studying local biodiversity effects of land use change, exploitation of natural resources, pollution, invasive species, and climate change.