Acting with Style: Towards Designer-centred Reinforcement Learning for the Video Games Industry

CHI 2021 Workshop on Reinforcement Learning for Humans, Computer, and Interaction (RL4HCI) |

Organized by Association for Computing Machinery (ACM)

In recent years reinforcement learning (RL) techniques have been successful in solving complex problems, especially in video games. However, this rapid progress has not yet translated into mass adoption of RL techniques in the video games industry. We believe there isn’t enough focus on being able to specify not only what goal our agents achieve, but also how they achieve it and also how reinforcement learning techniques fit into pre-existing workflows and constraints. We offer three suggested methods to alleviate these problems: Using preference learning to specify agent styles, using Potential-based Reward Shaping to make combining multiple sources of reward more robust and using an automated reward ratio scheduler to allow designers to work at a more meaningful abstraction level. Finally, we present a set of questions that we as a research community should answer to make reinforcement learning more approachable by the widest audience of potential RL users.