Building Non-player Character Behaviors By Imitation Using Interactive Case-Based Reasoning

The creation of believable characters is one of the most challenging problems in the interactive entertainment industry. Although there are different tools available for designers and programmers to define the behavior of non-player characters, it remains a complex and error prone process that requires a high level of technical knowledge. Learning from Demonstration is a promising field that studies how to build intelligent agents that are able to replicate behaviors, learning from demonstration of human experts. This approach is interesting for developers who do not have a computer science background, alleviating the need of representing tasks and knowledge in a formal way. In this work we present an online and case-based reasoning agent that learns how to imitate real players of Pac-Man using an interactive approach in which both the human player and the computational agent take turns controlling the main character. In our previous work, the agent was in complete control of the learning process so it decided when to give up or regain control of the character. Now the system have been improved so the player can also regain control of the character and go back in time to correct improper behaviors manifested by the agent whenever they are detected. We also present an evaluation of the system performed by three professional video game designers, followed by the main insights we have gained.