Mike Booth, the Architect of Left 4 Dead’s AI Director, Explains Why It’s So Bloody Good

Mike Booth, the Architect of Left 4 Dead’s AI Director, Explains Why It’s So Bloody Good
Image source: Valve
To sign up for our daily newsletter covering the latest news, features and reviews, head HERE. For a running feed of all our stories, follow us on Twitter HERE. Or you can bookmark the Kotaku Australia homepage to visit whenever you need a news fix.

A decade ago, Left 4 Dead’s chaotic, cooperative take on zombies was released into the wild. One of the best multiplayer games ever created, part of its genius was an ‘AI Director’ system that ensured every game felt fresh and unpredictable for players.

The key architect of that system and a big part of Left 4 Dead’s development was Mike Booth who, following a stint at Electronic Arts, co-founded Turtle Rock Studios in 2002. A huge Counter-Strike fan, Booth approached Valve about working on that series, and Turtle Rock subsequently did so for several years. After Turtle Rock was acquired by Valve in 2008, becoming Valve South, Booth eventually left for the equally titanic Blizzard. Since 2016 he has been at another tech giant: Facebook’s Oculus department, where he now specialises in social VR.

Kotaku UK: How satisfied are you with Left 4 Dead looking back now, compared to at the time of release? Would you change anything about it?

Mike Booth: I’m very proud of what we created with Left 4 Dead, then and now.

With the original Left 4 Dead, our game design approach was to keep things as simple as possible, but no simpler. In other words, there’s nothing in the game that didn’t need to be there. When you found a weapon, it was very clear what it did — a shotgun, an assault rifle, a hunting rifle with a scope. The design of the zombies were similar — the massive Tank, the Witch’s long claws, the Boomer’s massive gut. That kind of clarity really seemed to help players learn and understand the game, allowing them to focus more on the complexity of teamwork amidst the chaos.

I’m also pleased with the way our game constantly generated meaningful decisions and tactical options for the survivors. Should we shoot that gas can and fill the room with flames? Should we take up a position around that stationary minigun? Do we make a break for the saferoom now, or take it slow and steady? Should I throw a pipe bomb over there and attract the horde with its beeping? Is everyone ready to make the radio call, knowing all hell will break loose when we do? Each player had the opportunity to be a hero or to cause immense trouble for the whole team, and every decision mattered.

Image source: Valve

Are there any particular design choices you’d highlight?

The horde itself is full of intentional design decisions I remain proud of. Seeing a streetful of idly wandering zombies that clearly don’t realise you’re there creates such great decision moments and dramatic tension for the survivor team, knowing that a single noise could alert them all and cause total mayhem. The way that an individual zombie ‘wakes up’ and starts to look around is a clear visual cue for the survivors to be extra careful.

Even once a zombie notices a survivor, it takes a certain amount of time to focus in and start to run, making it very obvious what’s about to happen. The way the zombies lean into their run as they round corners, climb and clamber over everything, smash through doors, stumble back from a melee swing or explosion, and even the way they die — staggering and falling onto the environmental geometry — is due to intentionally designed algorithms and behaviours. It all matters, and is a core element of what makes the game fun to play.

People tend to think the zombie horde AI is simple, and they just “run at you”. The truth is that they were built on years of Counter-Strike bot navigation work, hundreds of motion captured animations from professional stunt people, and years of additional work to get their behaviours and interactions just right. We had a general design rule: “There can be nowhere the survivors can go where the zombies can’t reach them.” Making that statement true was a major challenge, and I’m very pleased with the result we achieved.

One specific creature design I’m particularly proud of is the Witch. I love how the sound of her sobbing in the distance is often more frightening than a rampaging Tank. I love how she tries to warn encroaching survivors to stay away by growling and slowly standing up. I love how incredibly dangerous she is once she has been set off. I particularly love that an encounter with the Witch can usually be entirely avoided, but somehow often isn’t.

What I’m most proud of, however, is how the game was able to successfully get strangers to work together and bond as a team. While it may now be common in modern games, at the time it was a big unknown to expect players who didn’t know each other to put themselves at risk to help an incapacitated teammate get to their feet, rescue someone dangling from a high rooftop, share pain pills, or to spend their only first aid kit to heal someone else.

Image source: Valve

Your initial cooperation with Valve sounds almost utopian — did that sort of developer collaboration happen often, and does it now that you’re in the world of big tech?

Looking back, it was a pretty ideal situation. We had built a great relationship with Valve through our Counter-Strike work over the years, and they were in a very strong position with the success of Half-Life 2, The Orange Box, and Steam. They used some of that strength to give us the flexibility and runway to really discover something new and make it great.

That kind of collaborative situation is very rare. And once you’ve had that experience, it’s very hard to work any other way.

Did you, and the team at Turtle Rock, have any anxiety about Valve buying into the project, in terms of expectation levels?

There’s always stress and worry in those situations, but we had built a lot of trust both ways and worked well together. I knew we would cede some creative control of the project to Valve, but that wasn’t a bad thing – they were world class game designers and we valued their contributions. I also knew the scope of the game had become a heavy lift for our small team at Turtle Rock Studios at the time and we needed Valve’s help to bring the experience across the finish line. For example, we were only building for the PC. If it wasn’t for Valve, there would not have been an Xbox version of the game. They helped improve the final game experience in many other ways as well.

What was the most stress-inducing period of development? Were there moments when elements of the game clicked into place, or where major ideas had to be scrapped?

At the very beginning of development, we built the game to be an unstructured city environment where the survivors could go anywhere. However, playtesting made it very clear that the moment-to-moment gameplay was so intense that players didn’t have the time or ability to plan new routes or agree where to go, and instead found the fastest path and went the same way every time.

While it was a hard decision to rework everything into a linear path, it was clearly the right one. With a single, clear route, the problem of procedural population and dynamic pacing became more understandable. Since we knew where the survivors were going, we could place ambushes ahead and behind them and control the intensity of the experience. This paved the way for the creation of the AI Director.

A bet we took early on was getting the zombie horde to climb over anything in their path. I was pretty confident we could do it based on my Counter Strike bot navigation work, and we went ahead with expensive professional motion capture sessions with Hollywood stunt people. When it all came together and we saw the first horde clamber over a chain link fence, we knew we were onto something really great.

Another “aha” moment was making zombies lean into their turn as they sprinted around corners. This was a tweak I made one afternoon. Since I knew the path the zombies were running via the pathfinding algorithm, I could use that to approximate a curve, and from that get the normal of the current curvature. Given that, I just tilted a sprinting zombie proportional to their speed and the path curvature. When we playtested that night and the first horde came sprinting around a corner down the street, it was a ‘wow’ moment for us. That lean, coupled with the motion capture sprint, made the horde seem much more real, believable, and intense.

Yet another challenging time in early Left 4 Dead development was when we were designing and building out the No Mercy campaign. That campaign was the first environment we built, and the feel of Left 4 Dead was incubated within it. We iterated on the gameplay and that environment hand in hand. However, for a long time No Mercy was one single, monolithic, map.

When we started to finalize production and build out final art, the giant map quickly became a severe performance problem. We simply could not reach the quality level we needed and keep the map as big as it was. At the same time, the pacing and duration of No Mercy was great, and we didn’t want to cut anything out.

The answer was clear — we had to chop the map into sections and load each in succession. That lead us to brainstorm the idea of a ‘loading room’ that loads the next section when you enter, which then evolved into our iconic saferoom with red lockable doors and a cache of supplies. From there, we landed on the rule of “if one survivor makes it, everyone continues.” So much drama in Left 4 Dead revolves around the survivors reaching the next saferoom, and we have the limitations of the Source Engine to thank for it! Another example of constraints driving creativity.

Image source: Valve

How risky did the AI Director feel as a design choice? Were there times when you worried it wouldn’t perform to the required standard?

Even though I had many years of building AI systems for games and simulations and had the codebase I had built for the Counter-Strike bots at hand, it was still a major calculated risk. It was a challenge to algorithmically populate the environment, but to also dynamically “direct” the pacing of the game was something I hadn’t seen before.

However, we had tried traditional methods of populating zombies, and nothing was really working. Players were just too good at memorising, which totally ruined the tension and turned the game into a race. If we were a larger studio, perhaps another solution would be to build many maps, or many, many variants of populations within a map. But even that wouldn’t give you truly enduring replayability.

Plus, I had an idea of how it could be done, and it was a really cool space to innovate in. I wanted to push the game design envelope and show that procedural content algorithms could not only work, but create a new kind of fun. So, I dove in and built it.

There were certainly difficult spots. Before the ‘director’ part of the AI Director was added, the experience was really uneven. While many games would be great fun, some games would be really, really easy with lots of empty streets and few zombies. Other times, a playthrough on the same map would be a brutal meatgrinder with tons of zombies everywhere. We had layers of randomness going on, which kept things unpredictable, but that also meant sometimes you ‘rolled’ a bad game. Randomness is, well, random. We couldn’t ship a game that was often really fun, but sometimes really not.

That was a pretty scary time. It wasn’t exactly clear how to solve the problem, and when the games were good, they were so very good! We didn’t want to lose those good game experiences. Part of the issue was tuning and experimentation, but it was clear we needed something that would guarantee we wouldn’t get the bad outlier games — either too easy or crushingly hard.

Once we landed on the idea of a system that could shut off the zombies and give the players a breather when they needed it, we were on the right track. I spent several months in spectator mode watching the rest of the dev team play through the game, monitoring AI Director parameters in realtime, and tweaking variables as we played. The final AI Director is a collection of clever algorithms, tricks, smoke and mirrors, and many, many hours of tuning and balancing.

Do you think that multiplayer narrative design has kicked on from L4D? Whether through incidental, environmental storytelling, or the use of AI systems for emergent play?

Emergent, procedural multiplayer experiences are still the exception rather than the rule in the AAA game space. In fact, there seems to be an strong inverse relationship between the size of a game studio and the amount of procedural algorithms used in their games. Major studios tend to hand-craft everything, whereas small indie studios build entire games procedurally.

There are exceptions to this rule, which are exciting to see. Games like Shadow of Mordor, Far Cry, Payday, and Vermintide come to mind.

With respect to using AI systems to create emergent gameplay, there have been some really exciting innovations in the last decade. Minecraft is a beautiful example of the kind of deeply rewarding, collaborative, and social gameplay that results from well designed, procedurally generated gameplay. Notch essentially turned an AI Director loose on the entire game world.

There are so many interesting indie games that make great use of procedural content and emergent game design it’s hard to talk about them all. Minecraft, Terraria, Don’t Starve, Spelunky, Rust, The Binding of Isaac, Crypt of the Necrodancer, FTL, Risk of Rain, Dead Cells, and Slay the Spire are a few I’ve played in the last handful of years. There’s clearly demand for the kind of unique experiences emergent procedural game systems deliver, and it’s great to see game developers continuing to explore here.

On the environmental narrative front, a great current example is what the Fortnite team is doing with their seasonal updates. What was that rocket launch for? What is deal the with these rifts? It really keeps you coming back to see what will happen next.

Image source: Valve

Similarly, how do you feel about the state of co-op? Squad systems in battle royales are meaning that co-operative play is still around, but bespoke small-team experiences like Left 4 Dead are thin on the ground.

I’m a big fan of battle royales. I play PUBG nightly, and have had a great time with H1Z1, Fortnite, Realm Royale, and most recently Call of Duty: Blackout. What is particularly ironic to me is during the development of Left 4 Dead I would often say things like “we want players to WANT to work together and move in the right direction without some kind of magical force field artificially pushing them forward.” While I still think that was a good design lens for Left 4 Dead, the explicitly shrinking zone in royale games creates a ton of anticipation and drama as well as great co-operative decision making.

Four player squads in battle royales are capturing some of the pacing and intensity we generated in Left 4 Dead with hordes of AI controlled enemies and the AI Director.

That said, one of my regrets has been not continuing to innovate in the cooperative/procedural game space. There’s so much more to explore here. We had only just started with Left 4 Dead.

Horror games are huge in the age of — but few are able to shock in emergent ways, let alone in multiplayer. How challenging was it to maintain adrenaline levels for players?

We discovered pretty early on that Left 4 Dead couldn’t be a creepy, scary, horror game. We certainly tried to keep some of that with elements like the distant sobbing of a Witch, or the unsettling music cues when you walk out into a dark alley. In general, that kind of spooky, creepy, scary is really hard to sustain and reliably recreate over repeated game sessions.

Instead we focused on giving players an “intense” experience. The ideal situation we wanted to happen often was the survivors just barely limping their way to the next saferoom, with a horde of zombies closing in — ideally a lone survivor, with the rest of the team cheering them on from spectator mode. Build the adrenaline up to a fever pitch, hold it there until it seems like they’re never going to make it, but then they do make it and the survivor team bonds over “OMG” moments they can’t believe they just experienced.

That we found we could reproduce reliably with the AI Director.

As for streaming — an emergent, cooperative game like Left 4 Dead designed to be streamed would be amazing. The AI Director would have enough information to choose dramatic camera angles for the stream, and maybe even narrate what is happening! Someone needs to build this.

Image source: Valve

Community toxicity is a major issue for multiplayer games currently; did this enter into your team’s thinking back in the mid-noughties?

Make no mistake, multiplayer games had toxicity issues ten years ago, too. We were very aware of the risks, and talked about the issue often. However, we really believed in the potential of co-operative game experiences, and there just weren’t very many at the time. We explicitly decided to take risks and rely on core game mechanics like reviving incapacitated players and rescuing survivors from locked closets.

We also made sure to include self-policing systems like vote kicking and invested in automatic survivor character vocalisations — “Pills here!” — so you didn’t always have to use a microphone. That said, it was all a calculated risk, and we didn’t know if random strangers would actually work together on the internet.

As it turns out, if you give people an exciting and clear goal (escape the zombies together), a well-structured activity (head to the exit), and clear ways to work together and help each other out, they usually will.

How do you feel about other games that have come out since 2008, riffing on Left 4 Dead – from twists like Payday to those that are more openly similar in structure, like Vermintide? Does that inspire pride, and any resentment at all?

It definitely inspires pride. That other developers not only appreciate the procedural design of our game, but then decide to invest their own time building upon our core ideas is quite an endorsement.

I’ve played both Payday and Vermintide and they each bring something new to the table. The thing about complex, emergent game systems is that they are inherently hard to reverse-engineer, and each of these games feel quite different and made different choices than we did. I think that’s a great thing, as they will likely discover new kinds of fun experiences that are unique to their game worlds.

I hope more developers build emergent co-operative game experiences so I can play them, too. There’s a game called GTFO coming out soon that looks interesting.

Any word on the grapevine about a third game in the series?

You’d have to ask Valve about that.

You Can Now 3D Print Working Game Controllers
This post originally appeared on Kotaku UK, bringing you original reporting, game culture and humour from the British isles.


  • Good read, I enjoyed that.

    Lovvved l4d back in the day, musta played it every night with 2 or 3 mates for about 3 years lol.
    Didn’t like l4d2 quite as much for some reason.
    Wish they’d do a l4d3, but this is valve we are talking about, very unlikely.

Log in to comment on this story!