Thesis Title: Social welfare in heterogeneous multi-agent systems

Data: 13 de Dezembro de 2021

Hora: 13:00 

Link de Zoom: https://videoconf-colibri.zoom.us/j/89779112698

Thesis Abstract:

The outcome of many of the most challenging problems humans face results from the combined decision-making of a large number of individuals. Examples abound -- from human cooperation, antibiotic overuse, climate change governance, or more recently, the adoption of voluntary measures related to the spread of SARS-CoV-2. When independent agents control different variables of a system, several problems emerge. In this work, we are mainly interested in the rise of cooperation problems that can have disastrous consequences. For example, when every country chooses not to reduce its CO2 emissions in the hope that other countries will, or when people choose to not get vaccinated in the hope that others will, a failure is inevitable and we face what is known as the tragedy of the commons. These complex multi-agent systems exhibit what are often called social dilemmas of cooperation. Because of the calamity of a defective solution and the difficulty to achieve cooperation, we are interested in studying their cooperation dynamics as well as developing solutions that can help agents escape destructive or catastrophic defection. The first part of the work is more descriptive and concerned with social simulation. The second part is more prescriptive and concerned with engineering solutions for the problem of achieving cooperation. We begin with the descriptive study and investigate the dynamics of cooperation in large populations facing threshold public goods games. We choose these games because they are representative of the aforementioned real-life problems. While several studies have shown that cooperation is hard to achieve in these games, we look into the additional impact that inequalities (and particularly wealth inequalities) can have on cooperation. Inequalities are ubiquitous in our world and it is important to understand what impact they can have on a society's aptitude to effectively cooperate and avoid collective catastrophes. We study the problem of inequalities both statically, from a game-theoretical perspective, and dynamically, using large populations of independent reinforcement learners. We find that wealth inequality does indeed significantly modify the game dynamics and strongly reduces cooperation within a population. This consequently increases the risk of disastrous outcomes. Our results can be instructive for individuals hoping to better understand social dynamics in these critical situations as well as for governments or entities that need to manage and regulate them. Interested in improving cooperation levels between reinforcement learners facing social dilemmas, we move to a prescriptive line of research. Our goal in the second part of this dissertation is to develop mechanisms that allow agents to avoid what we described earlier as the tragedy of the commons by facilitating the emergence of cooperation and hence social welfare. Here, for simplicity, we consider only 2-player games and look at inter-agent feedback i.e., remunerations and punishments to encourage cooperation between the agents. We develop a framework where agents use reinforcement learning methods to learn to punish or remunerate their opponents with the goal of resolving underlying conflicts in their environment and reaching higher cooperative levels. In environments presenting the agents with uncorrelated goals or even conflicting ones, the agents can use positive or negative rewards to align their opponent's goals with their own, hence increasing cooperation levels. We show that using such peer rewarding or punishing mechanisms allows independent agents to cooperate in situations where they otherwise would not. Our results give us insights into how the establishment of remunerations and punishments in our society may aid in achieving higher cooperation levels and increased public good.