Mary Lou Maher
University of Sydney
Creative behavior can be characterized as behavior that is novel, valuable, and surprising. Developing creativity behaviors in non-player characters is challenging because multiuser game environments are dynamic and changes are not predictable. Our ability to predefine task specific rules or rewards, or to define environment specific motivation signals, becomes more unlikely especially as these environments become more adaptive to the people that inhabit them. The development of more believable non-player characters will require computational processes that enable the character to focus attention on a relevant portion of the complex environment and to be curious about the changes in the environment. This talk will present computational models of curiosity, motivation, and attention focus that are the basis for a curious learning agent. These models are combined as a reward function for a reinforcement learning agent that is able to sense its environment and focus on novel and interesting changes in the environment that then serves as an intrinsic motivation for learning behavior policies. The result of deploying this agent in a non-player character is an embodied agent that expresses creative behavior.