An Adjectival Interface for procedural content generation

Abstract

In this thesis, a new interface for the generation of procedural content is proposed, in which the user describes the content that they wish to create by using adjectives. Procedural models are typically controlled by complex parameters and often require expert technical knowledge. Since people communicate with each other using language, an adjectival interface to the creation of procedural content is a natural step towards addressing the needs of non-technical and non-expert users. The key problem addressed is that of establishing a mapping between adjectival descriptors, and the parameters employed by procedural models. We show how this can be represented as a mapping between two multi-dimensional spaces, adjective space and parameter space, and approximate the mapping by applying novel function approximation techniques to points of correspondence between the two spaces. These corresponding point pairs are established through a training phase, in which random procedural content is generated and then described, allowing one to map from parameter space to adjective space. Since we ultimately seek a means of mapping from adjective space to parameter space, particle swarm optimisation is employed to select a point in parameter space that best matches any given point in adjective space. The overall result, is a system in which the user can specify adjectives that are then used to create appropriate procedural content, by mapping the adjectives to a suitable set of procedural parameters and employing the standard procedural technique using those parameters as inputs. In this way, none of the control offered by procedural modelling is sacrificed â although the adjectival interface is simpler, it can at any point be stripped away to reveal the standard procedural model and give users access to the full set of procedural parameters. As such, the adjectival interface can be used for rapid prototyping to create an approximation of the content desired, after which the procedural parameters can be used to fine-tune the result. The adjectival interface also serves as a means of intermediate bridging, affording users a more comfortable interface until they are fully conversant with the technicalities of the underlying procedural parameters. Finally, the adjectival interface is compared and contrasted to an interface that allows for direct specification of the procedural parameters. Through user experiments, it is found that the adjectival interface presented in this thesis is not only easier to use and understand, but also that it produces content which more accurately reflects usersâ intentions.