This paper presents the results of a study on the application of neural augmented topology evolution (NEAT) to game play automation in the challenging action RPG game Dark Souls. Unlike existing reinforcement learning or gameplay approaches, we evolve a neural network directly from raw pixel data, enabling game play without explicit game state information. To this end, we introduce Dark Souls API (DSAPI), a novel Python framework that leverages real-time computer vision techniques to extract important game metrics such as player and enemy health states. Using NEAT, the agent evolves an effective combat strategy to defeat the game’s first boss, the Asylum Demon, without any predefined actions or domain-specific heuristics. Experimental results show that the evolved agent achieves up to 35% success rate, demonstrating the feasibility of neural evolution in complex and visually sophisticated game scenarios. This study presents an interesting application of vision-based neural evolution, highlighting its potential for use in a variety of challenging game environments that lack direct API support or well-defined state representations.
