Modeling novel solutions to creative problem solving tasks with subjective observers
We propose a categorization of solution-centric evaluation metrics for a class of domain-independent AI challenge tasks known as MacGyver problems. Our definitions formally describe different classes of novel solutions for general creative problem solving tasks described in the MacGyver framework. Furthermore, inspired by existing theories of creativity, we extend the MacGyver problem formalism to incorporate subjective observers of problem solving tasks. By doing this, we explicitly model solutions to creative problem solving tasks as subjective evaluations based on the varying domain knowledge of observing agents. As an application of our extended formalism, we then illustrate how previous work on goaldriven conceptual blending represents a powerful form of human creativity whose creative solutions can be more formally described through our classes of novel solutions. Additionally, we conclude by highlighting strong connections between observer-oriented creative problem solving as described here and personalized procedural content generation in games.
