A Hybrid Approach to Procedural Dungeon Generation
Abstract
This thesis presents a novel approach to the Procedural Content Generation (PCG) of both maze and dungeon environments. The solution we propose in this thesis borrows techniques from both Procedural Content Generation via Machine Learning as well as Constructive PCG methods. The approach we take involves decomposing the problem of level generation into a series of stages which begins with the production of macro-level functional structures and ends with micro-level aesthetic details; specifically, we train a Deep Convolutional Neural Network to produce high-quality mazes, which in turn, are transformed into the rooms of larger dungeon levels using a constructive algorithm. For our dungeon’s micro-level details, we use a context-free grammar for the instantiation of interactable puzzle elements, and an n-gram model for decorating our dungeon's entrance rooms. This unique combination of methods successfully generates a large number of visually impressive game levels without compromising on any desirable PCG metrics such as speed, reliability, controllability, expressivity, or believability.