Autoverse: Evolving Symbolic Neural Cellular Automata Environments to Train Player Agents

作者：Sam Earle, Julian Togelius · 发表于：Proceedings of the Genetic and Evolutionary Computation Conference Companion · 年份：2025 · DOI：10.1145/3712255.3726722 · 被引用次数：1 · 研究领域：Reinforcement Learning in Robotics、Artificial Intelligence in Games、Cellular Automata and Applications

We introduce autoverse, a symbolically parameterizable Neural Cellular Automata (NCA)-based game engine for single-player grid-worlds. Users specify a set of spatial rewrite rules, which are sufficient for implementing a broad swath of the types of game environments (e.g. mazes, dungeons, sokoban puzzles) that currently serve as popular testbeds for Reinforcement Learning (RL) agents. These are compiled to a convolutions (i.e. an NCA), allowing for environments to be parallelized on the GPU, thereby drastically accelerating RL training. Meanwhile, rules and levels are encoded as binary strings, and applying mutations to these strings results in new, admissible environments, making autoverse a natural fit for evolutionary PCG.