Effective talk: AI for Ads on mobile games
Year of Talk: 2019
Video Link: Beyond NPCs: AI for Game Operations
The main purpose that the presenter and their team wanted from AI was its ability to assist people or do certain tasks faster and better than humans. This is so that the workers can concentrate on more important tasks and have a more impactful effect on the company.
Mobile Marketing:
Some explanation of how they do mobile marketing is necessary to understand the use of AI.
When trying to sell advertisements for a game an option is to communicate with a mediator who will contact with various Ad networks to sell slots on the game. The price is based on clicks per thousand. It will slowly go down if no Ad networks pick up the spots until one is chosen. Zynga (The company the presenter works for) mentioned that around that time they would work with between 40-50 different Ad networks. This had to happen in real-time for the users to receive their ad quickly. Which meant that the team believed there were efficiency gains for more ads.
The initial system to solve this efficiency problem was called Zoe. It starts by grabbing the data from the past 24 hours to be processed and aggregated for later use. After other processes are done on the data it's passed on to the Ad network configurations which are finally sold to Ad agencies. Any data acquired from the Ad agencies is collected and processed back into the Zoe system.
As part of the processing in Zoe, a wide range of scenarios are simulated based on the data and predetermined values. This allows Zoe to determine which combination of price and Ad agency might give the best value for the company. At times it was also tested with different variations of price being used. A caveat that is talked about is the fact that the distinct levers that the team tested worked independently and would affect everything in different ways and amounts. Even though during testing the values were never as accurate as initially intended there were many times that there were some improvements in amount of ads sold.
A similar problem is with User acquisition, specifically on how to properly optimize the amount of money being spent on ads and how to acquire more users. If you can make it cost less to get more users it helps with the amount of people playing and spending on the games. This can still be complicated as there can be various ways that people start playing the game and you might not have as much information as you want/need. A useful idea is to try and quantify how much different types of players are worth and how much they would bring to the game.
Propensity Modeling:
One of the simpler AI systems is propensity modeling. This model is created with historical data that has been classified and attempts to predict which classification the players might fall in. This gives it a probability wave on which bucket new data will likely fall in. It is a more common type of AI and a low-hanging fruit as it principally requires historical data that has been organized on a certain level into groupings. It is possible to make it even more granular if you want it to better sort future data.
Based on the historical data the model can determine how likely users are to stop playing the game or might not be having the best time. At this point, the AI can make certain decisions to help incentivize players to stay longer or to become more engaged with the game. It helps customize the game experience without a lot of extra effort from the side of the developers or Ad agencies.
The next step that can be done with the model is to help determine areas of improvement for the game. This comes down to the fact that it can determine what areas people are getting board the most with or find most challenging then give personalized experiences to keep them engaged. This personalized game helps better cater the game to a lot of players and makes them more motivated to support the game and devs.
Disclaimer:
Zynga is an extremely large company which means minor improvements can mean multiple millions of dollars in new revenue with minimum costs. Smaller companies might not be able to afford it and wouldn't see as many gains as a company like Zynga does.
It is important to understand that it can be slightly different for each game/product as there can be different reasons why people play the game. Or if AI-supported AD is something that would be useful for the game you are currently working on.
Comments