Teaching Embodied Agents through Human-Machine Collaborations

Sebastian Risi, IT University of Copenhagen

Abstract: In this talk I will give an overview of our work on evolving
and training agents to solve reinforcement learning problems. Given
their explorative nature, evolutionary algorithms in particular enable
new ways for humans and machines to collaborate to create complex
agent behaviours. I will show how casual users can create NPC
controllers for the Super Mario or Minecraft game without any domain
or programming knowledge. Additionally, I will demonstrate how human
insights can also be leveraged during the search for more cognitive
behaviors because of our ability to more easily distinguish between
adaptive and non-adaptive solutions. Finally, I will present our
recent work on combining generative adversarial networks (GANs) with
interactive evolutionary computation, which allows casual users to
more easily evolve specific target artefacts.