The variety of personas allows us to assess the generated content from multiple perspectives. To test the latter, artificial agents acting as play personas from prior work (Holmgard et al., 2018) are used to test the generated levels. Moreover, since constructive generators do not test the final result, the rules used in the different furnishers must ensure that the level can be completed but also viable for different playstyles. While creators only place walls or floor tiles and use tried-and-tested algorithms popular in rogue-like dungeon generation (Shaker et al., 2016), furnishers must account for the interactions and dynamics of different game objects. This allows for different combinations of creators and furnishers and can also work with manually created architectures (furnished automatically) or vice versa. The novelty of this approach is the use and analysis of a two-step process for generating the level’s architecture first, using a standalone layout creator, and distributing the game objects on that architecture based on game-specific rules using a furnisher.
The fast generation afforded by constructive approaches allows the game to create new levels with minimal lag even on a mobile device, for which MiniDungeons 2 is intended. This paper compares several constructive generation approaches that produce levels for the rogue-like puzzle game MiniDungeons 2 (Holmgard et al., 2016). Flappy Bird (dotGears 2013), Doodle Jump (Lima Sky 2009), Downwell (Moppin 2015), Polytopia (Midjiwan AB 2016), and Temple Run (Imangi Studios 2011) are just a few of many examples. Many mobile games use procedural content generation (PCG) to quickly generate content-with varying degrees of success. A better understanding of the properties of different families of generators as well as how they can be combined could help advance this situation.
However, choosing the right algorithm for the design constraints seems to be an art rather than a science. This allows games to quickly create endless variations to game-play by generating maps as in Minecraft (Mojang 2011), weapons as in Borderlands (Gearbox 2009) or NPCs as in Skyrim (Bethesda 2011) in real-time. Such generators are computationally lightweight since they do not evaluate their generated output. Unlike generate-and-test processes, constructive generators do not evaluate and re-generate output for example, cellular automata and grammars can be used for constructive generation, as well as more freeform approaches such as diggers (Shaker et al., 2016). , or machine learning (Summerville et al., 2018), level generation in published games is mostly carried out via constructive algorithms. While research on level generation focuses on level generators based on stochastic search (Togelius et al., 2011), constraint solving (Smith and Mateas, 2011 Smith
0 kommentar(er)
