Intrinsic Motivation in Computational Creativity Applied to Videogames
Abstract
Computational creativity (CC) seeks to endow artificial systems with creativity.
Although human creativity is known to be substantially driven by
intrinsic motivation (IM), most CC systems are extrinsically motivated. This
restricts their actual and perceived creativity and autonomy, and consequently
their benefit to people. In this thesis, we demonstrate, via theoretical arguments
and through applications in videogame AI, that computational intrinsic
reward and models of IM can advance core CC goals.
We introduce a definition of IM to contextualise related work. Via two
systematic reviews, we develop typologies of the benefits and applications of
intrinsic reward and IM models in CC and game AI. Our reviews highlight
that related work is limited to few reward types and motivations, and we thus
investigate the usage of empowerment, a little studied, information-theoretic
intrinsic reward, in two novel models applied to game AI.
We define coupled empowerment maximisation (CEM), a social IM model,
to enable general co-creative agents that support or challenge their partner
through emergent behaviours. Via two qualitative, observational vignette
studies on a custom-made videogame, we explore CEM’s ability to drive
general and believable companion and adversary non-player characters which
respond creatively to changes in their abilities and the game world.
We moreover propose to leverage intrinsic reward to estimate people’s
experience of interactive artefacts in an autonomous fashion. We instantiate
this proposal in empowerment-based player experience prediction (EBPXP)
and apply it to videogame procedural content generation. By analysing think-aloud
data from an experiential vignette study on a dedicated game, we
identify several experiences that EBPXP could predict.
Our typologies serve as inspiration and reference for CC and game AI
researchers to harness the benefits of IM in their work. Our new models can
increase the generality, autonomy and creativity of next-generation videogame
AI, and of CC systems in other domains.
Authors
Guckelsberger, ChristianCollections
- Theses [4099]