Analysis and implementation of multiagent deep reinforcement learning algorithms for natural disaster monitoring with swarms of drones

Abstract

Natural disasters are the cause of a great amount of deaths and economic loss every
year. The rapid and efficient deployment of mitigation and relief work is of great value but
can be hampered by the lack of information in real time about the situation. Technological
progress has allowed for a gradual improvement of this process of information retrieval.
Automatic detection of wildfires and floods from images is now a reality. However, the
real-time update of natural disaster maps still requires the physical presence of sensors.
A natural disaster surveillance strategy that is gaining traction in recent years uses
swarms of drones to acquire images of the environment. The cost reduction of this architecture
depends on the automation of swarm cooperative navigation.
The goal of this Master’s thesis is to build upon the recent advancements in the use of
Deep Reinforcement Learning techniques for swarm navigation to improve them and apply
them to a wider variety of situations.
In particular, we have developed a platform for the simulation and training of swarms of
drones and a series of multiagent algorithms. Lastly, we have trained deep neural networks
with and without memory in simulations of both wildfires and floods, following several
different strategies with varying levels of training decentralization, and we have studied the
potential benefits of each approach.
The final practical outcome is a series of trained neural networks which can be loaded
onboard the drones of the swarm to control their navigation in a decentralized manner.