Intentional computational level design

Abstract

The procedural generation of levels and content in video games is a challenging AI problem. Often such generation relies on an intelligent way of evaluating the content being generated so that constraints are satisfied and/or objectives maximized. In this work, we address the problem of creating levels that are not only playable but also revolve around specific mechanics in the game. We use constrained evolutionary algorithms and quality-diversity algorithms to generate small sections of Super Mario Bros levels called scenes, using three different simulation approaches: Limited Agents, Punishing Model, and Mechanics Dimensions. All three approaches are able to create scenes that give opportunity for a player to encounter or use targeted mechanics with different properties. We conclude by discussing the advantages and disadvantages of each approach and compare them to each other.

Procedural Content Generation through Quality Diversity

“…Expressivity analysis [30] is defined as the analysis of the output in terms of styles and variety of artifacts generated by the chosen approach, which can highlight biases of the generator towards specific types of content. While all [19].…”

“…In most of the previous work [14], [15], [16], [17], infeasible individuals are not considered in terms of their diversity. However, in some cases the diversity is always maintained regardless of constraint satisfaction [18], [19].…”

“…Constrained Surprise Search (CSS): Similar to constrained novelty search, this QD algorithm uses unexpectedness as the diversity measure for evaluating feasible individuals in a FI-2pop GA. As in the above variant, CSS is built on the FI-2Pop genetic algorithm, where the infeasible population minimizes the distance from feasibility and the feasible population maximizes surprise through surprise search [9], [22]. 6) Constrained MAP-Elites (CME): Combining the constraint-satisfaction capabilities of FI-2Pop with MAP-Elites, constrained MAP-Elites (CME) [18], [19] maintains two populations for each cell of the archive, one containing infeasible individuals and one containing feasible individuals. As in the original formulation of FI-2Pop [21], the feasible population maximizes the predefined fitness, while the infeasible population minimizes the distance from feasibility.…”

“…Moreover, QD algorithms which explicitly optimize for behavioral distance explicitly attempt to increase the expressivity of the generator by searching under-explored niches. Figure 1 shows an example of the covered space from the work of Khalifa et al [19] on generating scenes in Super Mario Bros (Nintendo, 1985), elaborated in Section IV. The figure shows that no levels are generated for some areas of the search space because the divergence characterizations chosen are dependent on each other.…”

“…Constrained MAP-Elites was used to generate small sections of levels (scenes) for Super Mario Bros [19]. The algorithm tries to find the simplest playable scenes, using an A* agent to check for playability.…”