“Not Bad For A Human” – The Quest for AI Game Designers

Disclosure: This video/blog was originally delivered by Tommy Thompson as a talk at the 2015 GameCity festival in Nottingham, UK. We’ve reproduced the talk for wider distribution with some embellishment of existing points.  Call it the ‘directors cut’, if you will.


Intelligent design is the cornerstone of the gaming sector, through which a variety of interesting and novel experiences emerge.  Design itself can be considered to be a number of functional and non-functional components or concepts that have matured over time.  I often associate good game design with a well-cooked meal.  It often takes time, attention to detail, a dash of common-sense and more often than not, the willingness to learn from failure.  As many game designers themselves will tell you, your first game won’t be great.  Making a game that looks, feels and plays well is a highly time-consuming process, largely because it’s not an easy problem to solve.

Despite or perhaps due to the the difficulty in creating well-designed games, it’s one area of games culture that people love to talk about.  Video game journalists and critics (myself included, given this series and format of delivery) pick apart the aspects of a game that make it resonate with both themselves and the larger player audience.  This comes despite the difficulty there can be in expressing your understanding of game design.  For many it’s the subtleties of a particular design that help reinforce a strong personal attachment, despite the fact it is very difficult for them to elucidate.  This is a common task I give to aspiring game developers who I have the fortune of teaching: just what is it about that game you love so much that brings you back to it repeatedly?

The underlying issue that sits within these discussions – ranging from the typical consumer to aspiring game developer and seasoned critics and creators – is that we often recognise good game design when we see it.  This is often reflected not just in sales, but the subsequent body of media that follows it: dissecting the minutia of what makes a game so satisfying, with designers subsequently invited to talk about their philosophies or ideas and how their games are a reflection of their practice.  This is arguably more common in indie game development; given that the final product is a reflection of the creative vision of a very small number of developers.

Super Mario Maker (Nintendo, 2015) lets you create your own Mario levels, but what makes for a fun gameplay experience?
Super Mario Maker (Nintendo, 2015) lets you create your own Mario levels, but what makes for a fun gameplay experience?

However, these ideas are not necessarily easy to translate into your own creative works.  A contemporary example would be to consider the community behind Nintendo’s level-editing software Super Mario Maker.  While players are fully aware of what constitutes a `Mario Level’ – thanks to over 30 years of continued releases – this doesn’t necessarily mean everyone can translate that knowledge into a fun level of their own.  Ultimately game design is something that is largely subjective, problem-specific and subject to a significant body of refinement.  It is for these reasons and more, that this proves really attractive to artificial intelligence research.

AI-Driven Game Design

Before we get into the real challenges of artificial intelligence (AI)-driven game design and the approaches taken, let’s look at what makes this so interesting from a researchers perspective. Much of this is driven from one of the core problems of game design: the notion of subjectivity.

AI systems are designed largely to be specialists at a given problem, with particular tasks defined and- more importantly – a series of metrics that help to quantify how well they perform that problem.  While research now exists that explores how well AI systems can become more general in nature – to the point they can solve any problem you give – the larger issue is how to quantify their performance with respect to that problem.  In many problem areas that AI is typically deployed, such as logistics and robot control, the metrics that dictate performance can often be defined pretty easily.  While we may have difficulty establishing whether we have achieved those metrics – in the case of say robotics due to issues with sensors – we can build definitions of ‘success’ and ‘failure’ that drive the decision making algorithms.

When dealing with game design, or any other artistic expression, beauty is in the eye of the beholder.  While we can perhaps quantify functional and logical aspects of a game artefact we cannot formally quantify the most aesthetic qualities of the AI-crafted game content.  Put simply, while we can quantify whether a level can be completed, or if a generated weapon is functional we cannot easily define numerically whether that level is fun or if that weapon is interesting to use.  These issues are arising at a time when game developers are seriously questioning what we mean when we talk about games.  The purely functional and logical design principles that have permeated since the the days of the Atari 2600 and the NES are now being challenged by a desire to be more aesthetically-driven.  Games such as ProteusGone Home and Everyone’s Gone to the Rapture still express logical components (which is to be expected of a computer program), but the gameplay experience is more focussed on eliciting emotional responses from players than satisfying any core gameplay loop (Cook and Smith, 2015).  This leads us into two specific areas of AI research and development: procedural content generation and computational creativity.

Procedural content generation (PCG) is formally defined as the “algorithmic generation of game content with limited or no human contribution” (Togelius et al., 2013).  This is still rather vague and terribly academic in its definition, but it can be condensed quite simply as “something that makes something” (Cook, 2014): a piece of software that can create an artefact that is useful within the context of the game that it is built within.  This requires some sort of computer algorithm to guide that process, but also an element of randomness attached to it, ensuring that it creates a variety of different content along the way.  The content the algorithm creates can vary wildly, with examples found in contemporary games including game textures (Tiny Wings), weapons (Borderlands), levels (Diablo III, Spelunky), experiences (Left 4 Dead) to even the composition of music (Killer Instinct [2013] and even The Legend of Zelda: Ocarina of Time) (Collins, 2009).

This idea of automated creation of content extends beyond just the creation of artefacts, which leads us to the world of computational creativity (CC) research: an endeavour to understand, replicate and enhance the human creative process.  This relies on an understanding of not only computer science and artificial intelligence but also behavioural psychology and philosophy. This leads to exciting innovations in the creations of songs, stories, poetry and even jokes!  This also leads to research in automated game design (AGD), which is focussed on creating entire games from scratch!

Approaches and Successes

There are a number of challenges faced when aiming to explore PCG, not to mention AGD research.  This comes down to a number of issues in how we interpret the problem as well as what we want our algorithm to do:

  • Theory- or data-driven? – How do we make intelligent decisions within the generative systems?  We need to use some metrics to help guide the content that we create.  Sometimes this is done by establishing ideas of good design beforehand that we then encode into the system (Dahlskog and Togelius, 2012) or we might grab data real-time by paying attention to what players are actually doing (Hastings et al., 2009).
  • Interactive or simulation? – Another important issue to consider is whether players have means to interact with that system?  It’s one thing to gather data about the player to tune the system, but what if the player can actually manipulate it?  Even more interesting, is whether the player is made aware that the generator exists and what they can do to manipulate it?

We see these ideas explored in a variety of research projects – and even published video games!  We take a look at some notable examples from the last few years.

The Mario Level Generation Competition

The Mario AI competition ran between 2009 and 2012, with an initial focus on the creation of AI that could play a the clone Infinite Mario written by Markus Person (Notch).  However, the competition expanded in 2010 with the introduction of a level generation track, where contestants were challenged with creating PCG systems that could create interesting levels.  This is by no-means an easy task, with a variety of different approaches forming throughout the period.  Each of the methods taken in the competition needed to adopt an element of player modelling, in that it would use data modelled from watching the player complete a sample level.  This sample data would then influence the level generator such that it creates levels that in some fashion reflected this data.  This can lead to some crazy results as shown in the video below, when the player is in fact an AI bot which has been optimised for the game.

[embedyt] http://www.youtube.com/watch?v=V06nEHw70b4[/embedyt]


Galactic Arms Race

On the other side of the research spectrum, we see the release of Galactic Arms Race (GAR): an indie game developed by a team of researchers and aspiring games programmers at the University of Central Florida.  Whereas the Mario competition was focussed on the creation of levels, GAR is built to create particle weapons in real-time using a machine learning algorithm which adapts to player data over time.  In a similar fashion to the Mario AI competition, this game relies on in-game data from the players experiences to influence its activity.  By paying attention to what weapons the player likes, it creates new ones that are offshoots, mutations or hybrids of existing ones in the game.  Think Borderlands only this time the weapon generation system is actually paying attention to the guns you like to use.

[embedyt] http://www.youtube.com/watch?v=MOKNoZSAGGQ[/embedyt]


ANGELINA

Throughout these examples we’re looking at AI systems building pieces of content as part of an existing game, but what if the system were to make the game itself?  ANGELINA is but one of many automated game design engines which are focussed on the creation of entire games from scratch.  Developed by Michael Cook during his PhD research at Goldsmith’s University in London, ANGELINA has worked through a number of iterations over the years ranging from classic 2D style games to the creation of fully playable games in the Unity3D engine.  This resulted in an interesting example of attitudes towards game design and development when two different games developed by ANGELINA were submitted to the Ludum Dare game jam.  While one was submitted with full disclosure that it was developed by an algorithm, the other was presented as the results of a human designer.  The responses from the Ludum Dare community based on the assumption of who (or what) made the game proved interesting to observe.  You can see some a Let’s Play video by Mike himself for the disclosed game made by ANGELINA entitled ‘To That Sect’.

[embedyt] http://www.youtube.com/watch?v=mY9BK8KW5G4[/embedyt]


Tanagra

Between the works of ANGELINA and the more rigid generative projects in the likes of Mario and Galactic Arms Race is research in mixed-initiative systems: AI systems that allow for the adoption of human input as part of the generative process.  While the first approaches were reliant on human-data, they never explicitly allowed players to intervene.  Tanagra, developed by Gillian Smith at the Center for Games and Playable Media at UC Santa Cruz, did exactly that.  Not only could the system develop entire levels for a platforming game on its own, but could allow for designers to draw specific areas of the level or adjust the pacing or difficulty of a level with the AI building around it.

[embedyt] http://www.youtube.com/watch?v=JaMr4_nEiYo[/embedyt]


Sentient Sketchbook

The last example we showcase is Sentient Sketchbookanother mixed-initiative system that is reliant upon player feedback.  In this instance, players are given the opportunity to craft their own levels but instead of the AI being the dominant system, it acts more in a facilitative role.  Sentient Sketchbook offers suggestions of alternative level designs based upon their habits.  In the event a designer decides that they wish to use one of the systems ideas, it pays attention to the suggestion they selected and aims to learn why this proved interesting to the user.

Closing

We’ve barely scratched the surface of this wide-ranging area of research and we can be sure that there is still plenty more to accomplish.  The limits of what AI systems can create in games is as-yet undefined, but there is huge potential in this area that is continuing to mature and develop over time.  Who knows how long it will be before we don’t have to wait 2-3 years for a game to be made, but for an AI level designer to make games at the press of a button.  This author imagines it will be some time yet, but there will be plenty of interesting things to explore until that time comes.

References

Collins, K., 2009 An Introduction to Procedural Music in Video Games, Contemporary Music Review, 28:1, 5-15, DOI: 10.1080/07494460802663983

Enjoying AI and Games? Please support us on Patreon!
Tommy Thompson Written by:

Tommy is the writer and producer of AI and Games. He’s a senior lecturer in computer science and researcher in artificial intelligence with applications in video games. He’s also an indie video game developer with Table Flip Games. Because y’know… fella’s gotta keep himself busy.

Comments are closed.