SPADENet: Skill-based Player Action Decision and Evaluation for Card Games using Deep Neural Networks (Online Rummy as Case Study)

In Online multiplayer skill-based card games, players compete with each other intending to optimize their reward. They have to make many decisions quickly as the game progresses and new information is revealed. Representing and comprehending such intricate information, and then making an optimum decision in real-time, is an extremely challenging task that requires skill. We introduce SPADENet, a generic and improved multi-input deep neural network architecture designed to 1) comprehensively capture and combine the players’ hand along with the ever-evolving visible game state outside of their hand 2) seamlessly incorporate the intricacies of different variants of any given card game 3) determine the best decision for any skill-based card games. Considering Online Rummy as an example, we demonstrate how this architecture can be easily adapted for determining the optimum drop (or play) decision across different Online Rummy variants (such as Points, Pool, Deals, etc.) and at different game stages. We observed improvement in Test F1 score across all Rummy variants with the highest being for Pool-6P from 0.767 (existing best-performing method) to 0.908 (18% increase). The odds of favorable game outcomes also increased when players follow model recommendations. We also show how comparable performance can be achieved by applying transfer learning across different variants. We further conduct a quantitative analysis of game state features outside of players’ hands to comprehend their significance in influencing the decision to Drop or Play. With this architecture, player skill can be measured and Low Skilled Players can be identified for any skill-based card game. We can take appropriate measures to improve their gameplay and provide a personalized upskilling path.