Evoluindo redes neurais artificiais para a geração de NPCs verossímeis utilizando visão artificial.

Abstract

This work approaches the generation of autonomous behaviors in No Player Characters using artificial vision. For this, variations with different resolutions of an Artificial Visual Cortex are proposed, represented by an Artificial Neural Network that receives the visual information (pixels) and turns them into signals for the NPC movement engines. During training, an evolutionary process using Genetic Algorithm is applied to Neural Network in order to teach it to recognize the patterns of pixels coming from a virtual camera. As new training generations are being created, the NPC is expected to be able to identify the Player in environment and follow it without any previous sensory description using only their visual perceptions. The results show that, in experiment scenarios, the Neural Network was able to evolve using only the visual sensing of the environment, generating the satisfactory behavior in the task of finding the Player.