Mobile app and game developers can use on-device machine learning in their apps to increase user engagement and grow revenue. We worked with game developer HalfBrick to train and implement a custom model that personalized the user's in-game experience based on the player's skill level and session details, resulting in increased interactions with a rewarded video ad unit by 36%. In this blog post we'll walk through how HalfBrick implemented this functionality and share a codelab for you to try it yourself.
Every end-user or gamer is different, and personalizing an app purely based on hard-coded logic can quickly become complex and hard to scale. However, taking all relevant signals into account, game developers can train a Machine Learning model to create a personalized experience for each user. Once trained, they can “ask” the model to give the best recommendations for a particular user given a set of inputs (e.g. “Which item from the catalog should I offer to a user that has reached skill level ‘pro’, is on level 5, and has 1 life left?")
Custom ML models can be architected, tuned, and trained based on the inputs that you think are relevant to your decision, making the implementation specific to your use case and customers. If you are looking for a simpler personalization solution that doesn’t require you to train your own model and where the answer is unlikely to change multiple times per session (e.g. "What is the right frequency to offer a rewarded video to each user?") we recommend using Firebase Remote Config’s personalization feature.
We recently worked with game developer HalfBrick on optimizing player rewards within one of their most popular games, with over 500M downloads, Jetpack Joyride. In-between levels, players are presented with an option to obtain a digital good by watching a rewarded video ad. Our objective was to increase the conversion rate for this particular interaction. Before the study, the digital good offered in return for watching the ad was always selected at random. Using a custom ML model, we were able to personalize which reward to offer using inputs such as the gamer’s skill level, and information about the current session (e.g. why they died in the last round, what powerups they selected), and increased conversions by 36% in just our first iteration of the experiment. Once trained, the model runs fully on-device, so it doesn’t require network requests to a cloud service and there are no per-inference costs.
ML workflows largely anchor themselves on to the following components: A problem to solve, data to train the model, training of the model, model deployment, and client-side implementation. We have described the problem above, and will now cover the rest of the steps. You can also follow along with more detailed implementation steps in the codelab.
Collect the data necessary to train the model
HalfBrick collected 129 different signals to train the model. These were of different types, such as event timestamp, user skill levels, information about recent sessions, and other information about the user's gameplay style. This is where intuition around which factors could influence the outcome are very helpful (e.g. a user’s skill level, point balance, or choice of avatar may be useful when predicting which powerups they’d prefer).
We can collect training data easily by sending events to Google Analytics for Firebase and then exporting them to BigQuery. BigQuery is Google’s fully-managed, serverless data warehouse that enables scalable analysis over petabytes of data. In our scenario, it makes it easy to summarize and transform the data to get it ready for model training, as well as combine it with other data sources our app might be using
Train the model, evaluate its performance, and iterate
Training requires a model architecture (e.g. number of layers, types of layers, operations), data, and the actual training process. We picked a reference architecture implemented in TensorFlow, and iterated both on the data we used to train it and the architecture itself. In a production workflow we would retrain the model regularly with fresh data to ensure that it continues to behave optimally.
In our codelab we have abstracted away the process of refining the model architecture, so you can still implement this workflow without having much of a background in ML.
Deploy the latest version of the model
Traditional ML models that run in the cloud can be easily updated, just like you would a web application or service. With on-device ML we need a strategy to get the latest model onto the user’s device. The first version can be bundled with the app, so that when the user first installs it there is a version ready to go. Subsequent updates are made by deploying the models to Firebase ML. With a combination of Remote Config and Firebase ML you can direct users to the right model name (e.g.
active_model = my_model_v1.0.3), which will then automatically download the latest available model.
Run the model on the device
Once the model is deployed on the device, we use the TensorFlow Lite runtime to “make inferences” or ask questions of the model based on a set of inputs. You can think of this as asking, “What digital good should I show a user based on the following characteristics?” We configured our implementation so that a certain percentage of sessions are shown a random option, which will help us gather new data and keep the model fresh and updated.
Try it yourself!
We have created a codelab for you to easily follow and replicate these steps in your own app or game. In this implementation you will encounter a scenario similar to HalfBrick's, but we've reduced the number of input features to 6 to make it easier for you to replicate and learn from. You should start with these and iterate as you see fit.
We hope you can use this ML workflow to build smarter and more engaging apps for your users and grow your business!