Open-Ended Evolution for Minecraft Building Generation

Abstract

This paper proposes a procedural content generator which evolves Minecraft buildings according to an open-ended and intrinsic definition of novelty. To realize this goal we evaluate individuals' novelty in the latent space using a 3D autoencoder, and alternate between phases of exploration and transformation. During exploration the system evolves multiple populations of CPPNs through CPPN-NEAT and constrained novelty search in the latent space (defined by the current autoencoder). We apply a set of repair and constraint functions to ensure candidates adhere to basic structural rules and constraints during evolution. During transformation, we reshape the boundaries of the latent space to identify new interesting areas of the solution space by retraining the autoencoder with novel content. In this study we evaluate five different approaches for training the autoencoder during transformation and its impact on populations' quality and diversity during evolution. Our results show that by retraining the autoencoder we can achieve better open-ended complexity compared to a static model, which is further improved when retraining using larger datasets of individuals with diverse complexities.

Transformer in Reinforcement Learning for Decision-Making: A Survey

“…For example, Minecraft [171] is an open-ended game with a sandbox design. Current TransRL algorithms [80] take a revolutionary application in Minecraft, which is beneficial to drive AI to express creative behavior by building structures that adapt to the world around it [172]. Like Minecraft, XLand (including versions 1.0 and 2.0) [173] is a productive prob game that consists of a huge variety of prob tasks in complex worlds with embedded hidden production rules.…”

Hierarchically Composing Level Generators for the Creation of Complex Structures

“…While there are numerous approaches to PCG, many of them focus on generating simple levels, and are therefore difficult to generalise to more complex settings, such as modern games. Many methods use some form of search [56], often evolutionary-based algorithms, to maximise a specific objective function [2,4,10,13]. However, it is challenging to design an objective function that is optimisable while also incentivising the generation of complex structures [54].…”

“…an entire, coherent town instead of just a single building). Many submissions to this competition, however, focused on more hand-designed, rule-based [2] methods specifically designed for settlement generation, which may require much effort to generalise to other scenarios. This has led to work that focuses on automatically generating lower-level structures one would find in a settlement, such as buildings.…”

“…This has led to recent research that applies PCG to more complex titles [2,17,22,39], one notable example being Minecraft. Minecraft is a 3D, discrete voxel-based game, making it amenable to grid-based PCG methods originally School of Computer Science and Applied Mathematics, University of the Witwatersrand.…”

“…Other approaches generate levels that lack some functional requirements, such as being playable [17], which is a core part of PCG. Lastly, some methods generate relatively small and simple structures [2,22], limiting their applicability to complex games.…”