چكيده به لاتين
In the gaming industry, creating well-balanced games is one of the major challenges developers are currently facing. Usually, games that are not balanced will have a high churn rate and will suffer in terms of monetization. Hence, nowadays a trending research area is focused on establishing mechanisms to create automatic balance in an algorithmic way. So far, various studies have used different methods such as neural networks, genetic algorithms and procedural content generation. Usually, these methods face extensive challenges such as time constraints, human errors, hardware deficiency and a small amount of initial data which will result in outcomes with low accuracy. In this work, the possibility of using deep convolutional generative adversarial networks for creating balanced video game levels is studied. In addition, a mechanism is developed to measure the accuracy of the generative network. For this purpose, a platformer game has been originally designed and developed. Then, the levels are randomly created while adhering to a set of balance requirements. Those levels that can be solved with the help of an agent using reinforcement learning are given as input data to a generative adversarial network. Finally, the network automatically generates new balanced levels. At regular intervals during the training of the generative adversarial network, weights of the network are stored to be used for evaluation. Then, generated game levels are checked to see if they have the game’s minimum necessary requirements and also to see if they can be solved by a human player. The best performing network is then selected for the generation of new levels. In the evaluations, it is shown that the proposed method is capable of generating levels that are well-balanced with considerable speed and accuracy.
Keywords: Generative adversarial networks; dynamic difficulty adjustment; reinforcement learning; Procedural Content Generation
