Why games are ideal for AI training

9 mins
Development of AI through gaming.
  • For decades, AI advancements have been put to the test through games.
  • But games also offer an effective training ground for AI to learn and improve.
  • AI training through games like Gran Turismo, Minecraft, and StarCraft could lead to real-life applications, from autonomous driving to business management.
  • We speak with AI expert Kenneth Stanley, inventor of the evolutionary genetic NEAT algorithm often used in training gaming AI.

  • Elevate your gaming experience with a gaming VPN to minimize ping and lag.

Artificial intelligence once made headlines for beating the world’s best at chess and Checkers. Today, AI is being put to the challenge in video games like Gran Turismo and Starcraft.

But these games are not just a way of assessing just how intelligent the machines really are. Even more so, games offer valuable training to AI, allowing them to learn as they play. In this article, we’ll explore this dynamic and the potential real-world applications of AI advancements gained through gaming. 

A brief history of AI in gaming

The evolution of AI in gaming traces its origins back to Alan Turing’s concept of the “imitation game” in the 1950s. Turing’s bold inquiry, “Can machines think?” set the stage for a transformative journey into the realm of AI and gaming.

In his thought experiment, Turing proposed a test that would forever shape the landscape of AI. Through the imitation game, a human judge would engage in text-based conversations with both a human participant and a machine. Should the judge be unable to consistently differentiate between the human and the machine responses, the machine would be deemed to have passed the test, achieving a level of conversational proficiency akin to human intelligence.

Turing’s visionary concept laid the foundation for exploring the intersection of human cognition and machine capabilities. It challenged conventional notions of intelligence and sparked the pursuit of creating machines that could emulate human thought processes.

Computer scientist Arthur Samuel coined the term “machine learning” in 1959. Samuel’s innovative fusion of artificial intelligence and gaming materialized in a Checkers-playing program. This program was unlike any before it. Samuel’s creation could learn from its mistakes, progressively refining its gameplay and strategy through experience.

The Checkers-playing program marked a watershed moment, illustrating the potential for machines not just to follow programmed instructions, but to adapt and improve autonomously—a fundamental shift that laid the groundwork for the evolution of AI in gaming.

And the milestones in AI gaming didn’t end there. In 1997, IBM’s Deep Blue confronted humanity’s reigning chess champion, Garry Kasparov, in a historic six-game match. The outcome—Deep Blue’s victory—reshaped perceptions of AI’s capabilities. The program’s strategic prowess and ability to outwit a human grandmaster heralded a new era where machines could rival human intellect in the realm of strategy and tactics.

These early chapters of AI in gaming sowed the seeds for further advancements, setting the stage for the astonishing feats to come. From AlphaGo’s triumph in the complex game of Go to the creation of formidable virtual adversaries in modern video games, AI’s evolution continues to captivate and redefine the gaming landscape, pushing the boundaries of human-machine interaction and paving the way for a future where gaming and artificial intelligence are inextricably intertwined.

A timeline of the development of AI through gaming.

Games as the perfect AI playground

Prominent AI labs, including those at Sony, Google, and Microsoft, have developed techniques that enable computer programs to conquer intricate board games and immersive video games with unprecedented mastery.

Kenneth Stanley, the former team leader of OpenAI’s Open-Endedness Team (which helps develop self-learning AI that can adapt to new tasks and environments), introduced the NeuroEvolution of Augmenting Topologies (NEAT), a genetic algorithm, which means it’s inspired by mutations and crossovers found in biological evolution. Developers have been applying the algorithm to real-time calculations in video games like the educational NERO and iconic titles such as Mario Bros. and the Monopoly board game. NEAT’s dynamic neural network adapts to player actions while the game is running.

For Stanley, games present an ideal testing ground for algorithms like NEAT. “Unlike costly robotic hardware, games require fewer resources and allow for rapid AI experimentation without real-world risks,” he explains in an exclusive interview with ExpressVPN.

But games have served not only as tests for AI but also as valuable training grounds. “In some cases, the motivation is making the game better, but in most cases, it’s about making the AI better,” says Stanley. “Games act as a vehicle for enhancing AI capabilities.”

“Sometimes the motivation for incorporating AI is to make the game better, but in most cases, it’s about making the AI better. Games act as a vehicle for enhancing AI capabilities.”

While driving simulators aren’t geared for AI, Stanley notes, games like Gran Turismo can be played by AI—and fact training using Gran Turismo led to the creation of GT Sophy, an AI with potential applications in self-driving cars. “Another good example of an AI sandbox is Minecraft with its limitless possibilities,” he says. “It allows for an approximation to real-life scenarios. Games are the most sophisticated kind of simulator.”

The unique combination of a constrained, controlled environment and free creative space in games makes them great for trying out AI methods. This helps advance AI and offers useful insights for solving real-world problems.

Examples of AI learning through video games and their broader applications 

You might think you get a lot out of gaming (entertainment, excitement), but AI that play games have the potential to gain a lot more—intelligence that can be extended to other applications. Here are a few examples of how that is being done:

  • GT Sophy, Sony’s AI racing driver featured in Gran Turismo, is trained via reinforcement learning, racing virtual cars for countless hours. This hints at future applications in self-driving cars and drones.
  • Microsoft researchers are testing an AI that allows users to build their worlds in Minecraft using simple prompts rather than hours of manual clicks. This could aid those challenged by traditional gaming controls and lead to broader accessibility solutions.
  • AI playing the classic Q*bert game discovered and exploited a previously unknown bug that allowed it to rack up unlimited points. The AI was simply trying to find the best solution and inadvertently revealed the bug. It bodes well for evolutionary algorithms, which the AI was based on, in which AI are tweaked slightly to find a version that performs best.
  • Teaching AI to conquer StarCraft, a complex multiplayer strategy game, means training it to gain management skills. The game’s challenges mirror real tasks—decision-making, strategizing, and resource management. A victorious AI would show algorithms can master real-life tasks.
  • A researcher from the University of Cambridge created an AI agent that can control characters in the battle simulator Pokemon Showdown, in which teams of six Pokemon compete against others. The AI analyzes the teams based on the characters’ strengths and weaknesses, predicting the outcomes. This could inspire technologies capable of managing teams in uncertain environments such as war zones.

Why humans can still beat AI at games

Despite advancements in AI, there remain games where human expertise prevails. Games such as Settlers of Catan, Dungeons & Dragons (D&D), and Cards Against Humanity stand as prime examples of challenges that AI struggles with. Even in games like Gran Turismo, Pokémon, and Monopoly, human players can still outshine their AI counterparts.

“AI will likely eventually master all conventional games. But first, we should ask ourselves what we define as a game,” asserts Stanley. “When games approach real-life complexity—such as designing machines or building a rocket—an AI would get in trouble. If it involves real creativity and too many degrees of freedom, it becomes challenging for AI to outperform human players. But in the long, long run, we can’t even be sure of that.”

This uncertainty emerges as researchers pursue the development of Artificial General Intelligence (AGI)—AI that is able to perform any task as well as a human can—to bridge this gap. But the training methods remain a question mark. “At the moment, we don’t know how to teach AI to be truly creative and come up with something new no one thought of before. It does happen in the small constrained worlds of gaming, but the real world isn’t small and constrained,” Stanley explains. “AI lacks instinct, but to teach it that, we need to know first what instinct entails.”

“At the moment, we don’t know how to teach AI to be truly creative and come up with something new no one thought of before. AI lacks instinct, but to teach it that, we need to know first what instinct entails.”

AI’s current limitations stem from data availability and its capacity to tackle intricate, open-ended tasks. Stanley clarifies that amassing the voluminous data essential for AI training, coupled with constructing networks capable of assimilating this data, presents a formidable challenge. Furthermore, AI’s reliance on textual information and its struggle with non-verbal or ineffable aspects further complicate matters. Existing AI models also struggle to understand chronology, a key element in comprehending novelty and intricate processes.

A solution might be lurking within the gaming realm once again. Researchers posit that Dungeons & Dragons, a game known for its collaborative storytelling nature, could serve as an incubator for AGI. Beth Singler, a digital anthropologist at the University of Zurich, introduced the “Elf Ranger Test” as an alternative to the Turing Test. This test suggests that if AI can adeptly engage in D&D, it might be edging closer to achieving AGI status. 

The future of AI in gaming and beyond

Looking ahead, AI’s role in gaming is set to expand even further. One clear trend is the use of narrative, social, and educational AI to make gaming more immersive and realistic. Fortnite, for example, introduced bots to train new players and a matchmaking system to connect players with similar skills. AI strides made in games like Starcraft II and Dota 2 are making games more customizable, helping them adapt in real-time to suit individual players’ skills, preferences, and tactics.

“To embrace AI without fear, view it as a tool for amplifying human abilities.”

Beyond the world of gaming, AI holds the potential to contribute to solving intricate global predicaments like climate change and healthcare advancement. However, this promise is coupled with concerns about AI attaining emotions and cognition.

Amid these advancements, concerns linger about the possible dark outcomes of using AI. “Concerns range from practical problems like disinformation or people losing their jobs to Terminator-like scenarios like the end of civilization. No matter how unlikely, however, you still want to make sure there’s zero probability”, says Stanley. “Ultimately, everything is creativity. If we cross that barrier, the world would not look the same, and it’s hard to fathom what that world would be like. What we do know, is that the real source of human pleasure isn’t consumption. The joy of being human lies in self-expression and creativity. So how can we protect that?”

Addressing these concerns is pivotal to shaping a balanced future. Stanley advocates for embracing AI’s growth instead of retreating from it.“To embrace AI without fear, view it as a tool for amplifying human abilities,” he says.

Enhance your gaming experience

Whether battling interstellar demons, crafting strange new worlds from scratch, or taking your team to the playoffs, integrating a VPN like ExpressVPN can amplify your gaming experience.

With user-friendly applications tailored to a range of major platforms, ExpressVPN lowers ping, optimizes speed and stability, and protects against DDoS attacks and ISP throttling. Additionally, consider exploring the ExpressVPN app for compatible routers, which extends its protective embrace to all devices on your network, including gaming consoles that lack native VPN support. ExpressVPN frees your gaming, ensuring unrestricted access to your favorite games anywhere, to play with your friends across the world.


FAQ: About AI Gaming

Is there an AI Dungeons & Dragons game?
Are there any games about AI?
Has AI gone too far?
Should AI have rights?
How can I make an AI that plays games?
Can AI make games?