Implementation of a 2D A. I. Agent for nondeterministic games using Convolution Neural Network

摘要

Artificial Intelligence (A.I.) in video games is largely used in designing an agent, also known as Non-Playable Characters (NPCs), that mimics characteristics of a human player. Evidently, the field of A.I. agent-based game development is flourishing due to recent advancement in the field of deep learning. This study presents the role of different Convolution Neural Networks (CNN) employed in the development of A.I agent-based games and also proposes a model which is a variant of CNN. For the purpose of this study the game used is a basic jumping platformer game in which the game character, which is of the primary focus, is tasked with dodging obstacles by either jumping or ducking. The obstacles are generated randomly and include objects like cactuses and flying vultures. The model proposed, is a variant of CNN that fashions an AI agent that predicts the game character's movement, based on the current environment of the game, to mimic the performance of a human player. The performance of the model proposed results in an accuracy of 97.03 % on both the testing and training set.