Particle swarm optimization for procedural content generation in an endless platform game

Abstract

Given the ever-increasing competition in the digital gaming industry, induced by a market of an exponentially growing gamer population, the production of creative, coherent, and appealing games has become inherently more complex. Creating game content by hand is both costly and time-consuming. By automating or assisting programmers and designers in their tasks, the techniques of procedural content generation (PCG) for games may address these challenges. PCG is not new, being active for several decades. However, the more traditional (and popular) form of PCG is somehow limited. It relies on some basic techniques ranging from simple pseudo-random number generators, generative grammars, image filtering, and spatial algorithms. The most advanced forms of PCG may utilize the modeling and simulation of complex systems and techniques from Artificial Intelligence (AI). In this context, this paper proposes a novel PCG approach, which is based on an optimization algorithm, known as Particle Swarm Optimization (PSO). Our approach is tested on a 2D endless platform runner game. The game structure (based on the Godot Game Engine) and a previous solution using Genetic Algorithms (GA) were first proposed and evaluated in an earlier work. In this paper, besides proposing the novel solution using PSO, we performed a comparative evaluation between the novel and previous solutions. The fitness function, utilized by both algorithms, takes into account the game’s environment aesthetics, physics, and some rules associated with gameplay, so that the generated environments are both enjoyable and playable. Our experiments evaluated time viability for in-game real-time generation and convergence to high/stable fitness values. We discovered hyper parameter ranges that yielded viable solutions. In the end, PSO has proven to be better suited to the investigated PCG task than the GA, since it presented faster convergence time and higher fitness function values.