The Automatic Generation of Game Environments for the Purpose of Training Artificially Intelligent Agents

Abstract

With the rise of self-driving cars and humanoid robots, it
has become important to validate the performance of AI agents in
simulated environments. In particular, simulated agents need diverse
environments to evaluate their skills. This presents an opportunity to
use automated methods to generate training data. The purpose of this
study is to compare the effects of training AI agents on various mixtures
of algorithmically-generated and AI-generated environments under
various test conditions. Inside a simulated environment, AI agents were
trained using reinforcement learning on different mixtures of artificially
generated environments. The results show that the agent trained on a
mixture of AI-generated and algorithmically-generated levels performed
best, while the AI trained on purely AI generated levels performed worst.
These findings show that using data from a mixture of artificial sources
may improve the overall performance of trained AI agents when faced
with limited data availability.