Predição de dificuldade em jogos Match-Three utilizando Redes Neurais Convolucionais.

Abstract

Since the emergence of the Match-Three genre in 1994, new games with this theme have been created, reaching its peak in 2012 with the launch of Candy Crush Saga, wich utilized new concepts such as maps with different shapes and diverse game modes. With the increasing popularity of games with this style, the need for constant content updates has arisen. One of the key points of content generation is the difficulty evaluation, wich envolves a great amount of time and tests with different users. Due to these factors, the difficulty prediction in games has gained a big focus on the game development area, removing the need of testing the created content with real players. In the context of Match-Three games, one of the metrics used to estimate the difficulty is made through the average human performance, wich is calculated by verifying the percentage of the average map objectives completed. Currently, there is the possibility of predicting this performance through simulations of the game state, making this process slow and computationally costly. This work presents a solution to the problem of the need for multiple game simulations, using Convolutional Neural Networks. In this way, it is possible to classify the difficulty of a map by just observing its visual aspects and its objectives. At the end of this work, it was possible to achieve a model of Convolutional Neural Networks capable of predicting the difficulty of 75,6% of the maps correctly.