How to save billions in game development?

back to our blogs

How to save billions in game development?

Gaming will exceed $100 billion in revenues in 2017. The global industry has been outgrossing movies and music for years and pushing the boundaries of consumer tech for decades.  It’s the prime target market for future VR and AR technologies.  Then why, when it comes to AI and Machine Learning, is gaming stuck in the past?

In July 2016, I had a chance to attend, a wonderfully organised, top-tier event for AI in the games and creative industries. How could I resist? I’d see the cutting edge of development and learn about better, faster, higher, cheaper, more exciting, freakin’ awesome new ways to do AI. I was a dog who’d been promised the thigh bone of a titanosaur. Full of anticipation, I boarded the plane to Vienna.

Not long after the conference got under way, all I felt was disappointment. It turns out that titanosaurs are actually fossils and don’t have much meat on them. But as the hours dragged by, as I listened, as I got my own thoughts in order...

I started to rally.

Disappointment breeds opportunity.

How can I explain this? Imagine it’s 1883 and you’re Karl Benz. You’re all energised by the possibilities of two-stroke engines and decide that you and your rich, adventurous wife Bertha should go to Vienna to attend the Konferenz uber die Futur des Tranzportation (obviously I’m making this up). You’ll discover the latest in transportation: technologies that will free the city of its 100,000 horses and their estimated daily - as in daily - output of 2.5 million pounds of horse-sheize. But instead of seeing presentations on new propulsion systems, all you get is the same old guff on how to breed faster carriage horses.

“Sheize, always more sheize”, you say to Bertha. “Umm?” “Wake up, Bertha! … What if we could build a horseless, asheizeless carriage … a Motorwagen?”

That’s pretty much how I felt about in Vienna.

Hand-Crafting is wasteful.

Presentation after presentation told of new ways to employ hand-crafted behaviour trees (a folksy name for decision rules) to build AI in games.  But handcrafting complex behaviour in games is expensive, unreliable and intractable - period. Doing it in 2016 was just plain dumb.

Yet even today, handcrafted behaviour trees dominate game-character development. As video games progress towards realistic graphics and virtual reality, players increasingly expect characters in virtual worlds to behave naturally — not in the stodgy,  predictable, rutted patterns that result from handcrafting.

The barriers to natural behaviour are obvious. AI agents that use handcrafted behaviour trees:

  • Don’t learn and adapt as the environment evolves;
  • Can’t handle complex behaviour with high dimensionality;
  • Don’t work in a multi-agent environments;
  • Can’t handle uncertainty in the environment.

The one thing they can do is cost the industry $10–15 billion a year.

Where’s the solution?

Enter reinforcement learning (RL): one of the most active research topics in AI and machine learning (ML).

RL can optimise sequential decision making in complex, uncertain environments. Some may still believe the myth that RL is not ready for practical applications, especially where the state of the environment isn’t completely observable. They’re wrong: the powerful beast of RL is difficult - but no longer impossible - to tame.

DeepMind demonstrated this with AlphaGo.   And VocalIQ, where I worked with RL researcher Dr Dongho Kim, proved reinforcement learning could be tamed for commercial outcomes even in partially observable situations like human conversation. That’s why Apple acquired the company.

Now is building platforms that get around the problem of handcrafting complicated behaviour trees. Dongho and I, along with co-founder Aleksi Tukiainen and a growing team of top-flight researchers, are applying RL’s decision-making frameworks to complex,  multidimensional scenarios.  The resulting platforms will allow game companies to use game-play data to build AI-bots that are self-learning and autonomous, yet tuneable and controllable. In complex FPS or MMORPG games, for example, non-player characters (NPCs) will be able to engage and entertain with surprisingly complex behaviours.

For players, autonomous bots can improve the gaming experience by:

  • adding (tuneable) unpredictability to games to keep them from getting boring;
  • handling very sophisticated and complex decision making;
  • evolving as the game evolves.

For developers, machine learning can cut time to market - and save billions - by:

  • dramatically reducing NPC development time;
  • enabling faster cycle times through accelerated feedback and 24/7 testing at computer speeds;
  • doing rapid, robust, functional verification through simulated play;
  • optimising performance verification and load testing with repeatable, virtual stress tests;
  • assisting human testers in coordinated teams.

That’s how can bring gaming back to the future.

And games are just the start. In 1888, Bertha Benz used her husband’s invention to become the first person to drive a car over a long distance. Some hundred and thirty years later we’re looking to apply’s technologies to robots, drones - in fact any autonomous system that requires complex decision making. And that includes self-driving cars.

More on that later.

Join us to make AI that will change the world

join our team