Adaptive dynamic game balancing based on data mining of sequential patterns

Abstract

It is the responsibility of a game or game level designer to provide players with a balanced game, one that offers a satisfying level of challenge. This can be done using traditional game programs and artificial intelligence (Al) techniques but it is becoming increasingly common for researchers are using dynamic game balancing, which uses reinforcement learning and focuses on the movement of non-player characters, especially in scripted games. However, this is not suitable for all game genres, such as those that use mazes or require dynamic terrains. In this paper we propose to adjust the level of difficulty of a game by using data mined for sequential patterns that can be used to analyze a player's behaviors. Our method first mines individual gameplay data and then transforms it into a set of sequential patterns. This approach is tested here on a maze game. We capture the behavior of several player's and the parameters of the game environment such as the number of dead ends achieved, hammers used on every level, how quickly players achieve the goal, the size of the maze, number of monsters hit, and the number of walls broken on each level. Feedback from participants in our experiments was very positive as they found the games designed using the proposed approach to be both more interesting and more balanced. This proposed approach differs from existing rule base game Al algorithms in three ways: (1) the game levels are based on the past experience of the player; (2) the approach is data-driven; (3) the game levels are unique and are not predefined, making them more adaptive, which contributes to making the game more interesting and balanced.