IT’S BEHIND YOU, hovering just out of reach, clearly conscious of your every move. You swerve left and right, dodging its advance while collecting valuable orbs from between the columns. It looks like you temporarily lost your mind there. But just when you think you have your enemy cornered, they come around the next bend and you both get hit with a vengeance. You are in it.
The game ends and the standard screen shows, but there’s a twist immediately below it. Please rate the complexity of this. You only lasted a few seconds in this terrifying pursuit, but while your heart was still racing, you instinctively tapped “Hard.” The thing that follows you around will be less hostile next time.
A Helping Hand from AI!
At Unreal Fest 2022, machine learning experts Romain Trachel and Alexandre Peyrot of Eidos-Sherbrooke showcased the game I just described. The Environment Query System (EQS) is a tool in Unreal Engine that allows designers to feed spatial information into AI decision-making processes; it is utilised here alongside machine learning.
Typically, this is taken care of with the help of behaviour trees, which organise the variables and branching options in a hierarchical structure. However, in this demonstration, the AI behaviour is being controlled by a machine-learning model. As the AI’s sensory organs, Unreal EQS takes in data about its surroundings for the machine-learning model to process before making decisions about how to react.
Gameplay-wise, it’s a basic cat-and-mouse chase in which players collect orbs scattered throughout a map, but the top-down presentation and placeholder images make it sound scarier than it actually is. Just like Pac-Man, except the ghost now acts independently of the programme.
“So, for example, if a developer opted to activate a greater pursuit mode, the only thing to do is increase a reference value in the EQS tests,” write Trachel and Peyrot in an email. Since in real-world game productions it would be up to a game designer to select which game variables must be modified in order to change the difficulty, this method “truly has the potential to streamline the development workflow.”
The term “up to a game creator” highlights the core idea of this explanation. To fine-tune behaviour, designers, programmers, and other developers may need to go back and forth through a complex behaviour tree. In order to simulate complexity without delving into the branches of a behaviour tree, designers may find that tweaking a machine-learning model is a more manageable alternative. If designers can put that aside, they can concentrate on whether or if the AI adds a sense of difficulty and enjoyment to the game.
It’s not always the case that a more capable manager is also the best manager
The use of machine learning allows for the creation of a formidable adversary. That was demonstrated by both IBM’s Deep Blue and Google’s DeepMind AlphaStar. However, this isn’t always desired, not just because it increases the game’s difficulty but also because the AI’s specialised strategies may be incompatible with fun gameplay.
Trachel and Peyrot experimented with artificial intelligence across multiple game formats, including a “multi-output model” that learnt to forecast the player’s score (acquired by collecting orbs) and sever them. However, in this play style, the adversary frequently camped near the orbs. We didn’t produce these outcomes since playing against them was boring and tedious.
A FEATURED VIDEO
the most well-known
Illustration featuring a mashup of browsers and cyber security logos.
The Time Has Come to Remove Your Reliance on LastPass
NEWMAN, LILY HAY
Cartoon depicts a person lounging on a sofa while reading, surrounded by a pile of books and a riot of colour and design.
The 12 Books You Need to Read in 2022
A KNIBBS, KATE
Illustration of a person using a smartphone to snap a picture, rendered in a bright, geometric style.
Everybody Misuses Google Photos
D. MATT BURGESS
Conceptual drawing of working remotely, featuring a campfire and a laptop-shaped tent.
This is the start of the “work-from-anywhere war.”
Since the player must collect orbs in order to progress through the game (imagine if Pac-ghosts Man’s simply waited near the entrances to each corner of the map), camping out at the locations where orbs are located is a viable tactic.
It’s a bummer because it ruins the game for everyone. There’s no longer any thrilling pursuit for the player. In its place, the AI could set up an unexpected trap. In their own words, Trachel and Peyrot’s objective is “not to develop superhuman bots—that would not be enjoyable and engaging for a rookie player” but rather “to identify ways to embed machine learning into game AI techniques already used in production.”
Better AI-hungry players may find that uninteresting. Despite the fact that the enemies the player faces in the final game don’t use the machine-learning approaches demonstrated by Trachel and Peyrot, these methods are nonetheless useful for tweaking difficulty.
For nearly five years, Modl.ai cofounder and research director Julian Togelius has been using artificial intelligence to the task of testing video games. Bots are used by Modl.ai to search for graphical errors, world geometry issues, and unwinnable circumstances.
You can choose the breakdown condition of your choosing. After then, it essentially just runs. Togelius explains, “You send off a job, and it runs based on how much you want to explore.” We can also group them together and give you a report detailing problem areas, etc.
Read More-How to Save TikTok Drafts to Your Phone
Modl.ai uses machine learning to customise its testing bots for each game, however the feature is currently only available for individual games. Togelius says they are piloting the integration of deep learning to teach bots to behave consistently across games. Once implemented, Modl.ai’s bots will learn to mimic real gamers’ actions, which should help them spot bugs faster.
Machine Learning in Games Requires a Paradigm Shift in Engine Design.
Therefore, machine learning can be both a source of and a remedy for trouble. However, developers that wish to include ML into games have more challenges than just coming up with engaging and balanced challenges. The issues go much deeper, and may ultimately necessitate a radical rethink of game design.
One obstacle is poor performance. In order for machine learning to provide useful results, massive amounts of training data must be collected, which can only be done by repeatedly playing the game dozens or even hundreds of thousands of times (though bots can lighten the load, a tactic Trachel and Peyrot used in building their demo). Furthermore, the resulting model can be difficult to execute in real time after sufficient training data has been acquired.
“Yes, performance is certainly an issue, particularly with huge ML models that analyse frames for each tick of the game clock,” said Trachel and Peyrot in an email. “In our situation,” they write, “we deployed a tiny neural network that was only inferring at exact periods of the game to prevent performance difficulties.” The issue arises when trying to scale up to the massive open worlds that current players have come to demand.
According to Togelius, the way today’s gaming engines are designed makes the issue much worse. He explains that since gaming engines were not designed for machine learning, the process will be delayed. Unreal and Unity and all its ilk are basically terrible—anti-AI in so many ways, which is why we don’t see more fascinating modern AI in games.”