Co-creating Platformer Levels with Constrained Adversarial Networks

Given the success of deep generative models in many creative tasks, it is natural to ask how to best leverage them to support human designers. We study this problem in the context of mixed-initiative design of platformer levels, a paradigmatic co-creative task. This setting is especially challenging, because – like all functional content – platformer levels must satisfy complex validity constraints, like coherency and playability. We explore mixed-initiative interaction with constrained adversarial networks (CANs), a class of deep generative models that synthesize structures satisfying one or more validity constraints. As such, CANs can be used to complete user-supplied partial levels while retaining full control of the constraints to be applied. We go one step beyond, and consider the issue of customizing a pre-trained CAN to some target design task or to the designer’s preferences. We discuss how to achieve this by combining CANs with coactive learning, a very natural mixed-initiate interaction protocol that acquires the necessary supervision from the designer in a transparent manner. Finally, we illustrate how to extend coactive learning to acquire informative supervision in the form of interpretable constraints.