Evolution and Learning of Cooperation in a Competitive Environment

Presentation 1
Personal Page
Presentation 2
Project Status
Project Poster
Short Paper

This page documents the Computer Science Honours project for John Richter, a student at Rhodes University.


Modelling artificially intelligent agents in an environment where they are required to co-operate in order to survive, while competing for scarce resources.

Basically, much work has been done creating artificial intelligences which are competitive, as many modern games can attest to. Several have some semblance of co-operative behaviour with and between agents. What I propose to model is a situation where AI agents need each other, but don't want each other. In order to maximize the comfort of their living conditions, they must put up with the others, while planning (through alliances, trust and possible betrayal) politically to take out those that are the greatest threats.

To be honest, many social and political constructions (like alliances and trust) are going to be quite tricky to model. Looking at the problem at a high level, I consider economic game theory to be the chief resource to be considered

Regarding the technologies and algorithms usable for such a project, I propose to build each of the agents (about 4 or 5 total) around a neural net. The nets shall be trained using genetic algorithms and reinforcement learning. If it is at all feasible, an expert-system-like 'memory' would be nice to implement, but that is possibly beyond the scope of this project.