This project consisted in developing new AI models combining interactive reinforcement learning with deep learning.
Reinforcement Learning agents are able to autonomously learn how to act in an environment, but are limited in their ability to generalize knowledge to unvisited states. Deep Learning methods, in contrast, are efficient to learn complex representations from data. The combination of the two approaches has shown promising in recent results considered as milestone breakthrough in artificial inteligence research.
Our approach consisted in adapting the neural network models at stake in deep reinforcement learning to an interactive, creative context, where humans would teach agents desired behaviours. We developed exploration bonus methods and formal user controls that jointly improve collaboration between humans and agents. These were implemented and evaluated in our software called « co-explorer ».
Van Kerrebroeck, MSc Thesis (2018)