King’s director of AI Labs Sahar Asadi has been at the Candy Crush maker for five years. In that time, AI use has exploded.
Practically every Candy Crush Saga level is now created with the help of AI, says Asadi. Her role is to help King’s teams understand players better and personalise King’s games to their needs, including using AI to pre-test levels before release.
She explained some of this in her talk at GDC, but also told us in person last month about how the Candy Crush Saga maker is automating level management in its games, while retaining human oversight.
“Before we release a level, we let the bots play that level and we get insights – are there any shuffles? How difficult was this level? And there are many more metrics that give designers an insight into the gameplay – we run them thousands of times so we have good accuracy,” she tells us.
“Then the designers decide if that is an intended experience that they wanted – yes or no – and they go back and refine the level. We also have built a tool on top of this playtesting that does automatic tweaking by AI.”
“But again, the tweaking is an assistive tool where the designer determines the criteria for what is a good tweak or refinement for each level. It’s an assistive tool, right? Consider it like a co-pilot for coding, but this is a co-pilot for designing.”
Asadi says that the more that King’s designers use their AI copilots, the better the tech gets – and could lead to a kind of prompt-based approach at some point in the future.
“If we go a few steps ahead in the future, a designer can say ‘I want this type of experience, this type of level’ and ‘this is an initial sketch of this level’…and then we iterate on that, we do the playtesting, we score under the hood, and we come up with some suggestions for the designer. And the designer can spend time on refining the experience.”
But King is careful not to let the new tech run amok. There’s a process for resolving instances where the AI suggests things that designers want to ‘overrule’.
“If we say ‘here is the tweak, this is the best one to make’…designers can explain the reason why this is not what they want to release,” Asadi says. “By designers annotating and saying ‘this is a good example’ or ‘this is a bad example’, the system is automatically updated and over time it learns by itself.”
This all sounds rather scientific, but personalising the Candy Crush Saga experience to suit different types of players is actually intended to be a little fuzzy, says Asadi.
King’s AI and human designers weight certain experiences towards certain player profiles, but it’s not a tightly prescriptive process – you can learn more about players by deliberately leaving in a little bit of haziness, it seems.
“The whole idea with personalization is that you don’t have these clear cut rules, saying that if this is one type of player you push them here, or push them there,” she adds. “It becomes more blurry.”
