Metica Solution Series
8 min read
By
Andrew Maher
ML engineering
Dec 18, 2024
Have you ever heard of contextual multi-armed bandits? Compared to A/B testing, they're like going from Mario's basic running speed to Mario on a mushroom-boosted kart.
Contextual multi-armed bandits (contextual bandits for short) are cutting-edge. They build on the classic multi-armed bandit framework, but add contextual data—like player behavior or preferences—to make smarter decisions in real time.
This revolutionary approach is already transforming industries like personalized recommendations and ad targeting, and now it’s poised to revolutionize gaming. With advances in AI and machine learning, contextual bandits are emerging as the key solution for dynamic, personalized experiences—adding far more value than traditional A/B testing. If you’ve ever felt frustrated by the limitations of A/B testing, contextual bandits might just be the game-changer you’re looking for.
This article is yet another in the theme of how to deliver true personalization in gaming. Here we’ll explain where A/B testing falls short, how contextual bandits work, and why they’re the key to unlocking your game’s full potential. For our previous blogs on personalization, you can find them all here.
A Common A/B Scenario
A/B testing is the de facto approach to trial different game experiences – from offers, to ads, to new gameplay mechanics. But it makes the critical assumption that there is a single “best” experience for all players. Often, this is wrong. Players are unique: they come from different backgrounds and they play games for different reasons. So what do you do when traditional methods like A/B testing don’t make sense?
Consider the following hypothetical example: Jake is a developer working for Space Mining Wars, a recent and successful game with good traction, great gameplay, and an engaged player base. Space Mining Wars is monetizing well enough—but it contains a progression pack designed to increase second-time purchases that few players are buying.
Jake has noticed this. He develops a few variants of the progression packs to display and A/B tests one of them against the existing pack. The test is designed properly and reaches a large number of players. Unfortunately, after running for a month, it yields no conclusive results.
When Jake performs a deeper analysis of the test results, he learns something surprising. Although the new progression pack isn’t globally better than the previous one, it resonates strongly with players in the U.S. What’s more, he sees that the original pack is popular with players in Japan and South Korea. But there’s nothing he can do: the results aren’t “statistically significant” for the full player population. Jake sticks with the original progression pack. Second-time purchases remain low.
Where A/B Testing Falls Short
This type of issue sound familiar? The example above is common in the industry. Monetization is notoriously difficult. Although classical A/B tests are great for choosing a single globally best game experience, they don’t enable the kind of personalization needed to unlock real revenue opportunities. This isn’t the only area in which they fall short. A/B testing struggles in several key areas:
Player diversity: Different player subgroups often prefer different variants, but A/B tests aren’t designed to handle this.
Low traffic: Smaller player bases make it hard to achieve statistically significant results.
Multiple variants: Testing many options simultaneously is inefficient and time-consuming.
For games, where personalization is critical to monetization and retention, these limitations are major roadblocks.
Enter Contextual Bandits: An Alternative Approach
Contextual bandits take A/B testing to the next level. Instead of searching for a single “best” variant, they adapt in real time, using contextual data to show each player the option that’s most likely to resonate with them.
Think of it like this: Netflix uses contextual bandits to recommend the perfect show, Amazon optimizes its page layouts, and Spotify tailors audiobook suggestions. In gaming, the same approach can revolutionize everything from in-game offers to gameplay mechanics - this is what Metica has developed for its’ customers.
Here’s how it works:
Players are initially shown one of several variants, much like in an A/B test
Over time, an AI algorithm identifies patterns between player behavior, attributes, and preferences.
The bandit dynamically adjusts, showing each player the variant that works best for them—automatically.
How Contextual Bandits Overcome Problems A/B Testing Can’t
Let’s go back to Jake and the Space Mining Wars example. If he had deployed a contextual bandit, three things would have been different:
The bandit would have learned that U.S. players prefer the new progression pack, while players in Japan and South Korea favor the original. It would have personalized the experience for each player group.
Jake could have tested all his new designs, not just one. Other regions, like the UK or Ireland, might have preferred one of the unused options.
The bandit would have uncovered complex relationships in the data that Jake couldn’t have identified manually.
With contextual bandits, Jake wouldn’t have had to choose a single “best” pack. He could have delivered personalized experiences to every player—and increased second-time purchases significantly
It’s Not Only About Complexity But About Speed
If we develop on this further, imagine a world where generative AI is redefining the speed at which new variants are created. In this evolving landscape, contextual bandits are more critical than ever. Traditional bottlenecks, like designing and producing new variants, are rapidly disappearing. Generative AI can now churn out hundreds of tailored game experiences, offers, or mechanics in a fraction of the time it once took.
The real challenge isn’t generating the variants anymore—it’s testing them, adapting in real time, and finding the right fit for every player. This is where contextual bandits shine. Unlike traditional A/B tests, which test one or two variants at a time and require weeks to reach conclusions, contextual bandits can rapidly evaluate multiple options, learning dynamically from every interaction.
By pairing Generative AI with contextual bandits, you’re not just creating a lot of variants; you’re unlocking their potential. This synergy allows you to deliver highly personalized player experiences at scale, and extract maximum value from your AI-generated content. In a world where speed, personalization, and adaptability are key to winning players' loyalty, this combination isn’t just a nice-to-have—it’s a must-have.
Conclusion: Why Contextual Bandits Are the Future
We’re not trying to say that A/B tests don’t work or are obsolete—they still have their place. But they're insufficient for building the kind of bespoke and personalized games that players want. Contextual bandits are essential. They’re an automatic AI solution that can optimize games, save time, increase retention, and boost revenue. Using them alongside A/B testing is a win-win.
At Metica, we specialize in helping games harness the power of contextual bandits. Want to see what they can do for your game? Book a free call with us, and we’ll provide a detailed analysis to uncover the opportunities waiting to be unlocked.
Want to find out more?