How AI is Used to Model Players | AI and Games Newsletter 28/08/24
We revisit classic AI and Games episodes on Team Fortress 2, Battlefield 3, Tomb Raider and more!
The AI and Games Newsletter brings concise and informative discussion on artificial intelligence for video games each and every week. Plus summarising all of our content released across various channels, from YouTube videos to episodes of our podcast ‘Branching Factor’ and in-person events.
You can subscribe to and support AI and Games on Substack, with weekly editions appearing in your inbox. The newsletter is also shared with our audience on LinkedIn. If you’d like to sponsor this newsletter, and get your name in front of our audience of over 4000 readers, please visit our sponsorship enquiries page.
Hello all, and welcome to another edition of the
newsletter! Today I’m summarising a stack of writing that subscribers will have seen popping up on Substack over the past week. I’ve been continuing my quest to dig up and resurface old work that I produced years ago and has been lost as websites and publications died - sadly all too common nowadays in games media.
I figured why not group together a collection of articles once again around a particular theme. Given I did this earlier in the year with Planning AI. This time around, we’re taking a look at Player Modelling: the practice of studying player behaviour in (and out) of games, and how it can impact the games themselves.
But before we do that, we have the obligatory announcements, and headlines in the news worth discussing.
Announcements
Alrighty, let’s kick things off with the obligatory breakdown of everything happening related to AI and Games, be it future events, projects and a little thing we call the ‘AI and Games Conference’.
AI and Games Conference
As mentioned in our last newsletter, I’m running the first ever AI and Games Conference on Friday November 8th at Goldsmiths, University of London - a project conducted in collaboration with the UK non-profit Game AI Events CIC and
’s .Two big things to highlight this week:
Early bird ticket sales will close on September 13th. A nice shiny discount for everyone keen to be first in the door. Our ‘Industry’ and ‘Student/Indie’ passes are selling fast, so grab’em while you can.
Our open call for submissions closes next week on Friday September 6th. After which our advisory board will review all submissions. Accepted speakers will receive a free complimentary pass to the event and be assigned a mentor from the advisory board to support them in making the best presentation possible.
We’re super busy right now putting together a lot of the back-end logistics. But we have plenty more announcements to make, with all of our speakers and their talks receiving the spotlight over on the conference social media page as well as right here on the AI and Games Newsletter.
Quick Announcements
Right now I have a stack of stuff I’d love to announce, but it’s a little too soon! I’m not allowed to tell you about it yet, but in the meantime here’s what’s coming up:
Don’t forget to catch last weeks return of the Branching Factor podcast, as
and I sit down to discuss the origins of the AI and Games Conference and a lot of the thinking behind its organisational structure and future plans.I’ll be delivering a bumper programme at the NEXUS Games Conference in Dublin, Ireland on September 26th. Not just giving a talk that morning on AI in the games industry, but will also be hosting a panel on AI for game development later that day. They’ve really got me working for that post-show Guinness.
I’ll be home a mere two days from vacation before swinging by the 2024 iGGi Conference which runs on 11th and 12th September at the University of York. If you’re not familiar, iGGi (the EPSRC Centre for Doctoral Training in Intelligent Games and Game Intelligence) is the world’s largest PhD programme for games researchers and hosts an annual event showcasing the research their students generate. I’ve collaborated several times with iGGi in recent years and joined the industry advisory board earlier this year. Will try and report back on what I catch while I’m there.
Right after my impending vacation, I need to finish writing my talk for GamesIndustry.biz HR Summit event is running on September 18th 2024, where I will discuss the potentials and pitfalls of AI for HR in the industry.
Right now one of the my biggest day-to-day projects is writing up my ‘Large Language Models (LLMs) for Game Designers’ course I’ll be running with Gamaste on October 7th in Paris, and October 9th in Lyon.
AI in the News
What’s next for artists suing Stability AI and Midjourney [VentureBeat]
An interview with several of the artists involved in ongoing litigation against AI image and video generator platforms. The class action lawsuit raised against these companies has successfully proceeded to discovery stage - which is still a long way from completion, but nonetheless a significant step forward. This isn’t going to impact the ways things are going today, but is certainly one to watch closely, given this could be a landmark case that dictates the future of generative AI tools.Amazon Games boss says AI won’t take away work, because ‘we don’t really have acting’ [Videogames Chronicle]
As mentioned recently, the SAG-AFTRA strike for performers is ongoing - and will be a future topic on the newsletter. But I was intrigued to see a comment by Amazon Games boss Christoph Hartmann highlighting that AI could and should be taking away the “really the boring parts” of development. Wise words.Creativity is made, not generated [Procreate]
An interesting public statement released by Procreate, creators of the popular iOS graphics application. In which they highlight “we’re here for humans”. This harks back to my comments a few issues back about the court of public opinion on AI in consumer products. Clearly Procreate are keen to distance themselves from much of what is happening in the current generative AI hype cycle.
A Primer on Player Modelling in Videogames
So for this week I wanted to break away from the news stories because a) I am doing a big chunk of writing for upcoming issues and b) I’m about to go on a short vacation and really needed something I could put together quickly. However, it did allow me to focus on one of my background tasks I had on the old to-do list.
Y’see, while this newsletter has only been running this year, and the Substack was setup in 2023, I have been writing about AI in video games online for about a decade now. This arose given I was wanting to provide another means for people to engage with my YouTube videos. This was largely for two reasons:
I appreciate that for a subset of my audience, they’d rather read an article from me on a subject versus watching a YouTube video. Either because it’s a better medium for them to focus on, it’s easier to do while at work, and the small contingent of people who just can’t get past my Scottish twang (shame on you really….)
A lot of my audience are students and scholars, and critically for them having a source that is available for citation, that isn’t a YouTube video, is often of great value. I know this from experience that citing an online video is still largely frowned upon in academic circles - and often with good reason.
Sadly many of these were originally lost to time, given old websites fell by the wayside as they became difficult to manage, or in some cases were brought down by hackers and others such annoyances.
So for this weeks issue, I’ve re-released (and also removed Substack paywalls on) a stack of my old blogs all about Player Modelling. Naturally these are a little dated when compared to some of the more recent work happening in this space, but I hope it gives some context as to what this field is all about, and gives a primer on how this work has been conducted across numerous video games in previous years. Plus as we’ll see shortly, some of this is merely an academic study from afar, while others are applications adopted within actual commercial video game releases.
What is Player Modelling?
So before I get into listing all of the articles I’ve re-published, let’s talk about the subject matter itself. Player Modelling is a process of data analytics in which we gather data about the performance of a player when playing a video game. The data itself can be gathered in many forms:
Internal Game Data: Information gathered from within the game, this can range from data about the current in-game activity or overall world state (game modes, overall team scores etc.) to behaviour of individual players (actions taken, current health etc.). Also the data that is collected can be at varying levels of granularity and fidelity. We may see what’s happening on a frame-by-frame level of the game such as the movement of characters and the actions taken, but we may also have processed higher-level data about perceived states of the game (e.g. an observed behaviour at a team level, rather than just individual).
Game Ecosystem Data: You can often grab relevant information about a player from connected adjacent systems that will reinforce the data coming from inside the game. This might be from distribution platforms like Steam, Xbox Network, and PlayStation Network, or other applications connected to the broader gaming ecosystem like Discord.
Player-Generated Data: There is lots of interesting data that can be gathered from a player outside of a game by attaching sensors or simply point sensors at the player. This can include things like recording facial expressions or voice, or even attaching sensors to read muscle tension, body temperature, brain activity and more.
Player-Perception Data: We can gather information by asking players directly, typically through use of questionnaires and surveys both in-game and externally to give us their perception of the game or their activities within it. This is often useful given player perceptions can and often will contradict some of the internal game data, and that can then lead to some interesting observations.
As we’ll see shortly, the examples I have collated over the years as articles for
rely on internal game data and player perceptions. However, you will find largely academic research that utilised player-generated data - in fact I have worked on several projects of this ilk over the years.But what can you then do with all of this information? Well in truth there are a variety of applications that you can build by taking this data, and then training an AI system using it.
Learn about Player Habits: You can identify common trends among players, often by identifying commonalities among subsets of your player base. This can include overall behavioural trends throughout a game - preferring a particular approach to the game as whole - or strategies towards completing certain parts of the game. It can even given insights into how gameplay intersects with other facets of the experience, such as how different player types engaging with monetisation models.
React to Player Behaviour: As players are making decisions within the game, you can anticipate what how they may react to decision points they will encounter in future, and then act based upon it.
Generate Content Based on the Data: Given we have information about how the player exists within a game, we can then build content that takes this into consideration. So a level generator could build levels based on the perceived skill level of the player.
Reproducing Player Behaviour: Perhaps the most common example within commercial titles is taking the players behaviour and then cloning it, plus adapting it such that it works in different situations.
Player Modelling on AI and Games
So let’s dig into some of the articles I’ve reproduced here on
all about player modelling research and applications. As you’ll see, the range of ideas being explored, and some of them are pretty wild.Player Performance Analysis in Tomb Raider: Underworld
One of the earliest examples of this type of work. A team of researchers at ITU Copenhagen analysed data pulled down from Xbox Live of players at launch of the 2008 entry of the Lara Croft’s adventures.Forza’s Drivatar
Perhaps one of the most well known examples of player modelling is the Drivatar within the Forza series. A system that captures information about players race in various vehicles and conditions such that it tell an AI to race ‘just like you’ for other players to experience.Killer Instinct
An area ripe for further experimentation, the 2013 reboot of fighting franchise Killer Instinct introduced the ‘Shadow’ system: a mechanism to create clones (i.e. shadows) of how a player behaves when using a certain character, that could later be made available to fight against on the Xbox network.Team Fortress 2
Research conducted at MIT was able to denote that there is a statistical correlation between perceived status and performance in Team Fortress 2 with all of those silly hats you can buy in the marketplace.Battlefield 3
A community-driven research project in which researchers accumulated in-game data from EA’s popular first person shooter and analysed it against data accumulated from surveys. It highlighted not only that player behaviour evolves in the Battlefield franchise as you age, but also that players are more likely to focus team performance than individual results the older you are.
This is far from an exhaustive list of applications of this type of technology in games, given it has also appeared in other work I’ve discussed in other capacities. Such as the AlphaStar project to train grandmaster bots for StarCraft II. Plus these are reproductions of old(er) writing of mine, and as such are not reflective of current, contemporary innovations, but it helps give a taste of the things that can be achieved, plus it’s nice to have these episodes back up on the website once again!
Wrapping Up
So after me going ahead and making changes to the format and timing of future issues, I am now about to break that rule. Given we have three issues of the newsletter arriving back to back, largely while I am also away again on a mixture of work trip and family vacation. Here’s what you can expect on the AI and Games newsletter in the coming weeks:
September 4th is all about the Generative Design in Minecraft Competition, as I invite Dr Christoph Salge to come onto the newsletter and discuss this years results. Plus Christoph and fellow organiser Victoire Herve (who you may recall wrote our guest entry on the PCG workshop at FDG 2024) will be on the Branching Factor podcast for a deep dive discussion.
September 11th we have a special issue in which we will be talking about Little Learning Machines: an indie game built by the team Transitional Forms out in Canada, that is all about training AI characters called 'Animos' to solve a collection of fun little puzzles.
And with that, I’m calling it a day. Thanks for reading and I look forward to having you back here on AI and Games next week!