Machine Learning for Game Master Recommender.

摘要

Artificial Intelligence and specifically Machine Learning has been put to use for years in the context of virtualizing board games. However, most of the current uses are typically applied to representing individual players or in-game agents; resolving tasks such as path finding or strategy. In the genre of more complex Role Playing Games there is another entity other than the normal player, the Game Master (GM). The GM's role is to not only serve as controller of the player's in-game adversaries, but also to construct the world, story, and scenarios the players are engaged in. It is not the goal of the Game Master to defeat the players (though this certainly can happen), but to create a challenging and fun experience for the players to engage in within the game's rule set. It is the goal of this thesis to encapsulate the thought process of a GM when it comes to constructing the game regarding what types of scenarios to present to players next after a sequence of completed scenarios. This requires the very human process of reading players' current engagement and attitude towards the previous sequence of scenarios in order to maximize their continued enjoyment. While previous studies have investigated altering games in real time to increase fun, these have all focused on manipulating the traits of the games' Non Player Characters (NPCs) (such as speed or difficulty) or integrated this concept into procedurally generated level design for one player games. What is being proposed with the GM recommender system is capturing the human behavior of reading the moods of players and overall story design through scenarios to increase the "fun factor" of a group of players. It will act at the level referred to as the "meta-game" and determine what sort of scenarios that give the most enjoyment between combat, skill checks, and story components, as well as variables specific to each of those scenario types (such as difficulty).