Show_AC_Thesis_Muayyad_Khatib_Hero.jpg

Information and Spatial Navigation through reinforcement machine learning

Simulating pedestrian behaviours can be a very complex task considering the different levels in which these behaviours occur: At the lowest level, agents interact with the environment based on a set of rules; at the mid-level, agents engage in more intelligent decision making (e.g., preferring a certain route), and at the higher levels, agents decide what past information about the environment is relevant for them to perform their future actions. Across all these levels, however, the agents base their actions on the knowledge available to them about the environment. In this paper, the aim is to explore this linkage and understand the correlation between the spatial information given to an agent, and its ability to explore new environments and generalise the information across multiple sites. This study will focus on developing agents using machine learning techniques, training them on simple environments and examining their behaviours in new more complex ones. Multiple types of intelligent agents will be created, each with different method of vision, allowing us to compare how different informational levels affect behavioural patterns, by mapping their navigational routes and testing their ability to complete given tasks showing explorative
behaviours.

Artboard 1-100.jpg
Artboard 2-100.jpg
Artboard 5.png
Artboard 13.png
vlcsnap-2022-09-11-17h11m26s362.png
Artboard 7.png
Artboard 8.png
Artboard 9.png
Artboard 10.png
Artboard 6.png