Deep Generative Models for Procedural Content Generation
Supervisor: Prof Simon Colton
Research group(s): Game AIDeep Learning has provided a step change in the ability of AI systems to recognise, classify and predict things. In addition, approaches such as Generative Adversarial Networks and AutoEncoders have also led to some increasingly sophisticated generative systems which often produce high value, aesthetically pleasing and surprising images, text and music. In the context of fluidic games - where the line between designing a game and playing the game is blurred - we have been looking at style transfer techniques to enable easy and rapid visual customisation of videogame assets. To take this to the next level would involve the software taking on increased creative responsibility in the design process, in effect becoming a videogame artist. Projects in this area would involve solving technical issues, in particular with the scope and speed of neural generative methods, and also solving human-computer interface issues, such as how a game designer can co-create in a mixed initiative setting with an AI system.