By John Stephenson
Machine Learning & Deep Learning
According to Newzoo’s Global Games Market Report for the most recent quarter, the video game industry will reach a market value of $139 billion by the end of 2018. Combined across mobile, platform, and PC games, video games have reached 2.3 billion gamers worldwide. Over a quarter of the world’s population has played a video game this year. That makes gaming one of the most wide-reaching forms of entertainment ever.
Major titles in gaming can rake in billions of dollars for game development companies. For example, Grand Theft Auto V is the most profitable entertainment product of all time at $6 billion in revenue, eclipsing all movies, TV shows, and music in terms of total revenue. Grand Theft Auto’s success, and the success of other popular titles including mobile games like Angry Birds or Candy Crush, is largely based on how thoroughly the game can build a world, capture the player, and provide tens or hundreds of hours of unique playable content.
In this quest for more realistic worlds, captivating challenges, and unique content, video game development shops are increasingly turning to machine learning as a possible ally in game development. Machine learning algorithms can respond to a player’s actions dynamically. Whereas everything in contemporary video games must be hand scripted, a video game with a machine learning engine could react and change how the world, non-player characters (NPCs), or objects behave in real-time, based on the player’s actions and decisions.
What Machine Learning Can Do for Games
Why are game developers looking into machine learning? There are essentially two problems in game development that machine learning can address in various ways:
Playing the game against (or alongside) human players.
Helping build the game dynamically for players.
We’ll explore the potential solutions in each of these categories below, but generally, machine learning algorithms can offload a lot of the work that a human game developer currently needs to perform. Control of non-player characters and the building of unique environments could all be automated if we can develop reliable algorithms for them.
There’s definitely promise in machine learning for gaming, but we’re nowhere close to being ready yet. Epic Games CEO Tim Sweeney has said that “[Video game] AI is still in the dark ages.”
However, once machine learning matures to a level that can reliably be used in games, it could fundamentally change the gaming experience in many ways:
1. Algorithms Playing as NPCs
Right now, your opponents in a video game are pre-scripted NPCs, but a machine learning-based NPC could allow you to play against less-predictable foes. These foes could also adjust their difficulty level. As you learn to play the game, your enemies could get smarter.
Companies are already working on early applications of machine learning-based NPCs. SEED by EA trains NPCs by imitating top players. Its NPCs learn dynamic movements and actions, and using human players’ actions as the training data means the algorithm trains four times faster than reinforcement training alone.
Teachable NPCs are a non-trivial improvement for game development. Currently, game studios spend hundreds of man hours scripting NPCs. Not hard-coding NPCs could reduce the development cycle for a game significantly. From weeks down to hours.
2. Modeling Complex Systems
A machine learning algorithm’s strength is its ability to model complex systems. Video game developers are constantly trying to get games to be more immersive and realistic. Of course, modelling the real world is incredibly difficult, but a machine learning algorithm could help with predicting the downstream effects of a player’s actions or even modelling things the player can’t control, like the weather.
One current example of complex modelling currently in production is FIFA’s ultimate team mode. As you select your team of all-star football players, FIFA calculates a team chemistry score based on how much the personalities on your team might get along, or not. During games, team morale can dip if you’re losing or making small mistakes. It can also surge when the crowd cheers and you’re playing well. The changes in morale impact the players’ abilities, in-game. More mistakes come when morale is low, and skill shots and lucky breaks happen more frequently when your team is playing well together.
3. Making Games More Beautiful
Another component of making games more realistic is making them look beautiful. Game developers are also using machine learning on this front. In a video game, often things look good from afar, but when you move closer objects render poorly and become pixelated.
Microsoft is working with Nvidia on this problem. They’re using machine learning to enhance images and renderings dynamically. In real life, when you’re far from an object the details aren’t clear, but as you approach you can notice finer details. This dynamic rendering of finer details is a challenge that computer vision algorithms can help with.
4. More Realistic Interactions
Another major challenge in building a realistic virtual world is how players interact with friendly NPCs. In many games, you need to talk to scripted characters in order to complete your objectives. However, these conversations are limited in scope and usually follow on-screen prompts.
Using natural language processing could allow you to talk out loud to in-game characters and get real responses, much like talking to Siri, Alexa, or Google Assistant. In addition, games that incorporate VR haptics or imaging of the player could allow computer vision algorithms to detect body language and intentions, further enhancing the experience of interacting with NPCs.
5. Universe Creation on the Fly
One of the most promising applications of machine learning in game development is world creation on the fly. To date, some of the most popular video games are expansive open map games that allow you to explore a massive landscape. Those games require thousands of hours of developer and artist time to render. However, machine learning algorithms could help with pathfinding and world creation. An example is a game like No Man’s Sky with an infinite number of new worlds to discover, all generated on the fly as you explore.
6. More Engaging Mobile Games
Mobile games account for 50% of gaming revenue industry-wide. Games on your phone or tablet are easy to pick up and play when you have downtime, without the need for a dedicated console. In the past, mobile games have been limited in scope because your phone doesn’t have the processing power and graphics of a console or PC. However, those limitations are starting to change with AI chips in newest smartphones that add specialised processing power. Many of the benefits of machine learning discussed above will become available to mobile games as well as the hardware continues to improve, making mobile gaming more realistic, interactive, and immersive.
The Future of Machine Learning in Gaming
There are still major challenges facing machine learning applications in gaming. One major challenge is the lack of data to learn from. These algorithms will model complex systems and actions, and we don’t quite have good historical data on these complicated interactions. In addition, the machine learning algorithms developed for the gaming industry need to be foolproof. They can’t break the game or the player experience. This means the algorithms must be correct, but they also must be fast and seamless from the player’s perspective. Anything that slows or breaks the game takes the player out of immersion in the world the game has created.
That said, most major game development studios have teams researching, refining, and applying AI into their games. This is a challenge that many companies are working on because it presents such an exciting opportunity to expand video games into new horizons, giving players even more realistic experiences and more playable content.