Dynamic Difficulty Adjustment with a simplification ability using neuroevolution

Abstract

This paper describes a concept of a dynamic difficulty adjustment system that learns and adapts online to the style of the player’s in-game behavior based on the neuroevolution using additionally the base of catalogues. A neural network trained through evolutionary algorithms is used to achieve better adaptation. In this case, genetic algorithms alter the weights of the neural network. In addition to adaptation, the proposed method provides an opportunity to downgrade difficulty when it is necessary or respond to rapid changes in the player skill level. The test game in the genre of real-time first-person fighting is designed to validate the efficiency of the proposed model. Multilayer perceptron (MLP) architecture was selected as a neural network’s topology to achieve the maximum correlation between an approximation of the loss function and speed of the neural network. The article describes the advantages and limitations of the proposed concept in comparison with other approaches.