Game Action Modeling for Fine Grained Analyses of Player Behavior in Multi-player Card Games (Rummy as Case Study)

Abstract

We present a deep learning framework for game action modeling, which enables fine-grained analyses of player behavior. We develop CNN-based supervised models that effectively learn the critical game play decisions from skilled players, and use these models to assess player characteristics in the system, such as their retention, engagement, deposit buckets, etc. We show that with a carefully constructed input format, that efficiently represents the game state and history as a multi-dimensional image, along with a custom architecture the model learns the strategies of the game accurately. It is further enhanced with look-ahead achieved by self-play simulation to better estimate the game state, and this information is used in a new loss function. Next, we show that analyzing the players with these models as reference has immense benefit in understanding player potential in terms of engagement and revenue. We also use the model to understand the various contexts under which players tend to make mistakes, and use these insights to up-skill players.