The purpose of this thesis is to investigate whether deep reinforcement learning can be used in the process of norm emergence in multi-agent systems. Multi-agent systems are systems composed of autonomous individuals that interact with them to solve complex problems. Norm emergence regards the emergence of patterns of behaviours of individuals in a society and has been the subject of study in a variety of fields, including machine learning. On the other hand, deep reinforcement learning applies deep learning techniques such as neural networks to reinforcement learning, where individuals, called agents, interact with an environment and learn by taking actions resulting in rewards. The application of such a technique to study the behavioural patterns that emerge in societies can provide exciting results to understand better how societies function. Traditional norm emergence techniques involve a third party in the learning process that coordinates and enforce rules whenever a conflicting state takes place, limiting the autonomy of such agents. Deep reinforcement learning makes the process of learning such rules more natural. Here agents recognise erroneous states and decide to adopt a behaviour to avoid them. Such a process makes agents learn rules instead of imposing them by also coordinating the agents through a rewarding systemIn this work, we simulate a road intersection environment where agents can move and interact with it, here agents need to reach the other side of the intersection without crashing into each other, doing so results in a positive reward while crashing in a negative one. Over time, agents start cooperating to cross the intersection and develop behaviours to solve such a problem allowing them to obtain the maximum reward possible. Two kinds of deep reinforcement learning methods are implemented which differ from the use they make of policies(which actions to take in a given situation): On-policy methods, which closely resemble norm emergence, and Off-policy methods. The entire system is built in Unity, making the entire process similar to building a video game level. Overall, this thesis provides an interesting view on the application of deep reinforcement learning in multi-agent systems to study the emergence of behaviours in individuals by showing that such a method yields similar results to traditional techniques and exposes the strength and limitations of taking such approach in norm emergence and multi-agent systems, by also providing some examples on how to expand and take further such research.

Link: to dissertation Paper: https://drive.google.com/file/d/1V545XsBR2xLV0zlCasFwu_-leac5LMH1/view?usp=sharing