Abstract: GameTune allows developers to optimize their mobile game for each player in order to improve metrics like retention or lifetime value. This is done by setting up a question together with a set of answer alternatives from which GameTune chooses upon request in real-time. For example, a question concerning tutorial difficulty (easy/medium/hard) could be asked in the beginning of the first gaming session. Another question could concern the frequency of interstitial ads (low/medium/high), which is a reasonable thing to adjust in the beginning of every gaming session. The former example is naturally modelled by supervised learning whereas the second setting requires reinforcement learning. Uncertainties play a significant role in the modelling task, both in decision making and in player behaviour. In this talk I describe how we use Bayesian neural networks and distributional reinforcement learning to tackle these problems.
Speakers: Mikko Kemppainen, Senior Data Scientist, GameTune
Affiliation: Unity Technologies