Using reinforcement learning to improve human learning in arcade games

Main supervisor: Herman Kamper

Reinforcement learning (RL) is a branch of machine learning where an agent learns which actions lead to favourable outcomes through trial-and-error experience. In conventional RL, the aim is to get an agent to perform well in some complex decision-making task, e.g. learning to play an Atari game. This project will flip the goal. Instead, the goal of the RL agent will be to learn how to change an arcade game so a human player can become better at the game as quickly as possible.

As an example, imagine a machine trying to teach a human to play better Tetris. It might be that, for the human to improve, the game should become more difficult. Or maybe the player struggles with a certain combination of block types. In some cases, it might even be best to make the game easier, if the player is struggling so much that they are getting nothing right. To hand-engineer a machine to change the game in an appropriate way would be extremely hard. We therefore turn to RL, where the agent's reward is based on the human's performance.

The example above is for Tetris, but the project will first focus on simpler Atari games. You will gain experience with cutting-edge machine learning approaches, particularly in RL.