Welcome to part two of the Coda series demonstrating how machine learning is an integral part of what we do, and how it provides fantastic opportunities for game developers.
In part one, which you can read here, we looked at how Coda uses machine learning to predict the lifetime value of a game.
In part two we are going to look at another way that Coda harnesses machine learning and that is in the use of tagging. As you’ll discover it not only enables developers to see what’s trending, but also helps to guide them with choosing the fundamental elements of their own specific game that will make it special.
Balls Master perfectly tuned all its game elements to make it successful upon launch
Once again Muhammed Miah, ML Engineer at Coda, is on hand to explain some of the more technical elements of the process.
What does the gaming landscape look like? Understanding that is really important to us at Coda. We want our developers to have the best shot at making really big hits, and so we need to know what type of games are being made as well as what type of games people want to play.
Here I describe how we go about understanding the games that everyone is producing and which ones are becoming popular. As usual, we start off requiring data. We need as much information about the games from the App Store as possible. This will enable us to see what is popular, what works and what doesn’t. By automatically looking at the App Store continuously, we make sure that our developers can always be on top of the trend.
We were inspired with the idea of annotating all 200,000+ hypercasual games on the App Store, both new and old. This effort had actually already started at Coda, but manually, and was already helping us understand the components that made the most popular games successful. This was clearly important work, but for a human highly tedious, and given that it could be quite dull, there was risk of inaccuracies. It also turns out that humans would have to spend over a decade, at a comfortable pace of 50 games a day, to complete the entire App Store.
Since waiting a decade was not an option, we turned to artificial intelligence. This is exactly the type of work that machine learning algorithms were created for. It actually turns out that the manually tagged games themselves were one of the key pieces that allowed us to leapfrog into using AI in the first place, and we are very thankful for that. It provided, as we call it, training data.
What elements of games fit well together?
So Coda harnesses machine learning to analyse and categorise games, allowing them to be easily tracked and segmented. With this we are able to uncover a wealth of insights to power our Market Intelligence tool.
Helping Machines See
Our Market Intelligence tool uses computer vision technology to tag games that are on the App Store. It is able to tell, for example, whether a game is 2D or 3D or if it is cartoon-ish or a puzzle game. One of the biggest elements of games, its game mechanic (how you play it), is also inferred. We pay special attention to what the most successful games are using, and what is trending.
Here are some of the things we look for:
|Game Mechanic||IO, Idle, Puzzle, Rising Falling, Swerve or Tap Timing|
|Control||Drag & Drop, Hold & Release, Swipe or Tap|
|View||Back, Isometric, Side, Top|
|Style||Abstract, Cartoon, Low Poly, Pixel or Realistic|
|Theme||City, Nature, Sports|
The technology we use here is called deep learning. More specifically, we use a feed-forward convolutional neural network with several layers.
“Let’s start with the new games coming on to the App Store,” explains Muhammed. “The system grabs all of the screenshots of each game from the store and feeds them into the neural network. This means that it only has the raw image pixels to work with, as in the actual colours of the game, and it runs these through the network to deduce all of the different categories. Some categories are easier than others, and it uses probability to make informed guesses. We then label the game with those tags on our platform. This is an ongoing process and occurs literally every time a new game hits the App Store.”
The platform here shows the most popular tags for each game mechanic
Intelligence For Developers
Collecting, sorting and labelling games is one thing, but what can you then do with that information?
“You might already have an idea for a really cool game, or you might need some inspiration,” says Muhammed. “Would you not want to know how well that idea will do? Come to the Coda platform and browse through the Market Intelligence area. Say that you want to create a game with a certain artistic style and a certain game mechanic. You can input these into the tool and it will come back with a list of games that match what you had planned to make. You can see which of those games performed the best. I would recommend you delve further into the Market Intelligence tool to understand what exactly made those games successful and apply those principles to your own game.”
“An additional feature of this tool is that you can see which tags are trending,” adds Muhammed. “Have pixel art games started becoming popular again, or is everyone interested in school-based themes now? Perhaps the drag & drop game mechanic is falling out of favour? Our market trends feature will tell you.”
Trending Mechanics: Puzzle games have become more popular recently
“We also took the liberty to revamp the lookalike feature,” Muhammed mentions. “In the same way that Amazon and Netflix offer recommendations, you can see which games are similar to the one that you are exploring. These recommendations are now a lot more reliable.”
“Maybe at some point in the future we will use deep learning to actually play the games itself.”
If you would like to learn more about how Coda uses machine learning, have a look at our previous articles: here and here.