Sokol Murturi

Sokol Murturi

AI for game design: learning from designers

The demand for games and game content has seen significant growth over the past decade, and developers are struggling to keep up.  Games such as ‘The Last of Us’ and ‘Bioshock’ are brilliant examples of how well content- heavy story-driven games can be received, yet the production time for these sorts of games can be 3 or 4 years.  Meeting demand by hiring ever larger teams of programmers and artists is not always a solution, especially for smaller independent companies.


Games such as ‘No Man’s Sky’ take an altogether different approach, using procedural generation create a vast universe of content.  Research into AI- driven game design has shown that there is potential to push this much further, with generation of personalised content and even whole games. However, a major concern with AI driven design is the loss of authorial intent and control over the final game as experienced by the player.


For my PhD I would like to investigate how AI can help developers by learning to generate content in a similar fashion to the developers themselves.  I envision a framework based on reinforcement learning, where an AI can learn a design policy for some content domain (e.g., FPS maps or platformer levels) by observing human designers.  The AI would learn to take particular design actions in certain kinds of content states.  Recent research into reinforcement learning has shown it is a powerful framework for developing complex agent behaviours and I believe there is a lot of potential to apply this work to game design.


How would a human and artificial designer interact?  Assume that an AI has learned to design a specific kind of content, such as a house, by observing human designers at work.  A human designer could then partially develop some new content, and ask the AI to suggest some variations on it (see figure below), with both AI and human iterating on the design in a mixed-initiative interaction.  The AI could learn from feedback from both the human designer and playtesting.  As human feedback may not produce enough data for effective learning, the AI could perhaps extend this with data from simulated playtests.

 sokol diagram

Game design decisions are often made with an expectation of how the player will react, and I could also look at how player models could be incorporated into the AI designer.  In a reinforcement learning approach, the state could represent content+player, and the AI could learn to take design actions aimed a specific types of player.   Developers could use this framework to develop content targeted at an individual player's style.  Moreover, if the AI has learned something about how the human designer creates content, it can then be used live during the game to modify game elements in response to player interaction.  Developers could set up modular levels, giving the AI the ability to adapt certain areas with content generated specifically to match the player.


I think that by learning from designers and players, we can address the major concern at the heart of AI generated content.

Home institution: Goldsmiths

Supervisor: Dr Jeremy Gow

Ready to apply?

Once you have identified your potential supervisor, we would encourage you to contact them to discuss your research proposal.

Learn More