Learning Player Behaviors in Real Time Strategy Games from Real Data

Abstract

This paper illustrates our idea of learning and building player behavioral models in real time strategy (RTS) games from replay data by adopting a Case-Based Reasoning (CBR) approach. The proposed method analyzes and cleans the data in RTS games and converts the learned knowledge into a probabilistic model, i.e., a Dynamic Bayesian Network (DBN), for representation and predication of player behaviors. Each DBN is constructed as a case to represent a prototypical player’s behavior in the game, thus if more cases are constructed the simulation of different types of players in a multi-players game is made possible. Sixty sets of replay data of a prototypical player is chosen to test our idea, fifty cases for learning and ten cases for testing, and the experimental result is very promising.KeywordsCase-based Reasoning (CBR)Real Time Strategy (RTS) GamesDynamic Bayesian Network (DBN)Junction Tree