Navigation in Three-Dimensional Ecosystem Models

Abstract

As the area of deep learning advances, it has become possible to predict and simulate areas of the real world in a digital setting. One emerging field is the research regarding virtual twins. Virtual twins aim to replicate aspects of the real world so that actions can be analysed digitally without needing physical access to the location. This thesis aims to research if it is possible to model reinforcement learning agents to traverse such three dimensional environments that are based on real world data.

By using a combination of land-cover data and heightmaps, the ability to generate three-dimensional environments in the game engine Unity was explored. The agents were modelled as three-dimensional animals which had the possibility to consume resources, run and rotate. These agents were trained in one of the aforementioned generated three-dimensional Unity environments, using a simple image containing multiple channels to represent the state around it. It was also these channels that the agents got as their input, essentially acting as their perceptions. Results show that the agents were able to learn and survive in the environment. Simple population tests also showed that agents were able to move to new areas to find food if there were too many agents in close proximity to them, which would lead to depleted resources. Furthermore, other tests related to the environment were done, such as how an increase in sea level could impact the environment and the population of the agents.