Evolutionary algorithms for generating interesting fighting game character mechanics
Eirik Høgdahl Skjærseth, Harald Vinje · NORA - Norwegian Open Research Archives · 2020
Abstract: Procedural content generation (PCG) is the process of generating video game content through algorithms and has been used in the game industry for a long time. "Content" refers to elements in a game such as levels, terrain, or game rules. PCG has several benefits for the industry since computer-generated content can inspire human designers, and be used to create games with endless content. Research interest in the PCG field has grown over the last couple of decades, where PCG problems are often formulated as search problems. Evolutionary algorithms have frequently been applied as the search mechanism, with an objective defined by a desired content quality. The optimal quality in entertainment games can be formulated as the question "will a player enjoy this content?" This quality has been quantified by heuristics within different game genres, with success. A promising approach is based on evaluating content while it is played in a game, referred to as simulation-based evaluation. Constraint novelty search is an evolutionary algorithm that has emerged and shown promise in the field of PCG. The commonly used objective given by a fitness function is replaced by the objective of novelty alone. This thesis is a first attempt at using constraint novelty search, with constraints based on simulation-based evaluation, to generate character mechanics in a fighting game. Character mechanics refer to the technical design of characters, i.e., how they may behave in a game. Aesthetics like graphics and sound are not considered. A simple two-player fighting game will be used for this research, developed by the thesis authors. The game is based on design patterns seen in commercial fighting games. User studies are conducted to evaluate the generated character mechanics, based on the enjoyment of test subjects. The results show that the applied generation method can produce multiple characters that are perceived as more interesting than human-designed characters. However, characters of high quality are not generated consistently. The amount of test subjects is limited, such that further research is needed to verify our results. The proposed generation method can be applied to other fighting games on the conceptual level. However, the implemented system is dependent on technical aspects and heuristics of the game used for this thesis.