Application of ethical reinforcement learning to a resource gathering scenario

Abstract

[en] In this project we present an application of a formal framework for defining moral values to a multi-agent system simulation of a society facing a social dilemma. First, a description of the framework and the motivation and key concepts for the understanding of this project are explained. Then we describe the case study: A resource gathering scenario, where agents have to face a dilemma between being benevolent and helping others or not, which has an obvious impact in the survival rate of their society. We use a Python 3 framework for agent-based modelling, MESA, and describe its structure along with which classes will be used in this project. We will also describe the class design for the implementation of the project as well as any other design decision. Our goal is to successfully add a moral dimension to learning agents by modifying its learning process, through the usage of norms, in order to instill our desired moral values. The results are discussed and compared to what we expect to be the optimal performance of a society facing said dilemma. We are interested in measuring its cooperation, which impacts directly in its survival rate, with and without the application of moral values. An improvement is expected to be seen in those measures when moral values are applied. Last, further work and possible projects derived from this one are also discussed as well as possible improvements to this project.