Learning latent representations for controllable combinational creativity and game design
Anurag Sarkar · 2023
Latent variable models have been increasingly applied for performing a variety of creative applications, primarily in the domains of visual art and music. Such models learn continuous latent representations of data which are then utilized for generating novel artifacts via sampling and interpolation, as well as for performing various other creative tasks. However, despite a growing body of work surrounding procedural content generation via machine learning (PCGML), the use of deep latent models for similar applications in games remains underexplored. While defining and using a possibility space of an individual game is a well-established practice in automated game design and procedural content generation, learning possibility spaces such that they span a set of one or more given games is uncommon, and in general, the use of generative models to enable a broader range of creative applications has not been as widely adopted for game design. Thus, in this thesis, we study how deep latent variable models can be leveraged for various game design applications, in two broad directions. First, we investigate the use of learned latent spaces for developing controllable combinational creativity systems, focusing specifically on game blending. Combinational creativity is the branch of creativity that focuses on producing novel artifacts by recombining properties of existing ones. Game blending is a combinational creativity process referring to recombining the levels and/or mechanics of two or more games to generate a new game and has been proposed as a means of capturing the process by which designers often create new games by combining ideas from existing ones. In this part, we focus on using variational autoencoders (VAEs) for building systems for performing such game blending, building up to a novel combinational creativity framework that defines and generates blends as linear combinations of learned latent design spaces. Second, we focus on using learned latent representations to enable game and level design applications more broadly. This section thus focuses on using models trained on one or more games to enable creative ML applications and affordances for game design, similar to those seen in visual art and music. We refer to these using the umbrella term Game Design via Creative ML or GDCML. More specifically, this part of the thesis demonstrates the use of supervised methods and evolutionary algorithms to enable a range of game design applications in the form of level editing, level search and optimization, level layout generation and style transfer.--Author's abstract