GVGAI as a Tool for Game Design

Diego Perez-Liebana, Jialin Liu, Simon M. Lucas
Links: – –

The General Video Game AI (GVGAI) framework and competition have attracted many practitioners, researchers and students during the last couple of years. This benchmark proposes the challenge of creating agents that are able to play any game it’s given, even if it’s not known in advance, in the absence of any domain knowledge. This is proposed in different settings, from single and two-player planning problems to learning without a forward model. Besides, the competition presents another two tests for generality, namely the automatic generation of levels and rules for any game. After briefly introducing the framework and its different tracks, this tutorial will show a new feature recently added to the framework: the possibility of setting the parameters of a game described in VGDL and explore the different possibilities the game rules offer. We will describe how to make a VGDL game parametrizable, and how to define a game space for it. Finally, we will show how an evolutionary algorithm, the N-tuple Bandit EA in particular, is able to tune game parameters to make it either more balanced or more challenging.

A Tutorial on Procedural Content Generation via Machine Learning

Sam Snodgrass, Adam Summerville

Procedural content generation (PCG) is the creation of content through the use of algorithms. Recently, there has been a growing interest in PCG via Machine Learning (PCGML). In this tutorial, we will give an introduction to PCG and PCGML techniques that have been previously explored. We will then guide the participants through a PCGML project, from data collection, through learning, and finally generating new content.

This tutorial is aimed at anyone with interest in game AI, especially those interested in procedural generation. Participation in the project portion of the tutorial will require programming experience and a computer on which to complete the project. Experience with machine learning may be beneficial, but is not required.

This tutorial will be broken into two parts. First, we will give an introduction and overview of PCGML and recent techniques. Second, we will guide the participants through the implementation of two recent PCGML methods: multi-dimension Markov chains [SO16] and long short-term memory recurrent neural networks [SM16].

For the first part of the tutorial, we will introduce the area of procedural content generation. Then, we will discuss the use cases of PCG, such as automatic generation of content, mixed-initiative content generation, and data compression. Next, we will present several recent PCGML approaches and how they can be applied to the above use cases. Finally, we will mention some of the current open problems facing PCGML, including style transfer, small data-sets, and evaluation of content.

After the participants have been made familiar with the field, we will move onto the guided project. We will begin by introducing them to the video-game level corpus [SSMO], a repository that includes level and map data-sets for several classic games, including Super Mario Bros., The Legend of Zelda, and Lode Runner. We will take this opportunity to discuss the importance of choosing the proper data representation. Once introduced to the available datasets, we will present a more detailed look at two specific PCGML techniques: multi-dimensional Markov chains [SO16] and long short-term memory recurrent neural networks [SM16].

Finally, we will guide the participants through implementing one or both of these techniques, using levels from one of the data-sets included in the VGLC. We will demonstrate how our respective models are trained, and how to implement the training procedures. After the participants have trained models, we will discuss various sampling techniques that could be used to generate new content.

At the end of the tutorial, the participants should have an understanding of Procedural Content Generation as well as how machine learning techniques could be leveraged in that domain.

Sam Snodgrass is a fifth year Ph.D. candidate in the Computer Science department at Drexel University. He is a member of the AI and Games lab under the advisement of Santiago Ontañón. His area of research is Artificial Intelligence. Specifically, he is interested in how machine learning techniques could be leveraged for generating content in the domain of games. Sam has published several papers exploring this area at conferences, such as AIIDE and IJCAI, and has an article in TCIAIG. He has previously held research internship positions at the Naval Research Lab and Army Research Lab. Sam is looking to graduate in the next academic year, and is currently searching for post-doctoral and faculty positions.

Adam is third year Ph.D. student in the Computer Science department at the University of Cal- ifornia, Santa Cruz. He is a member of the Expressive Intelligence Studio under the advisement of Michael Mateas. His research interests are heavily focused on PCG, specifically, how machine learning can be used to learn from existing artifacts to be able to generate, analyze, and cri- tique. Adam has been on the programme committee for the PCG and EXAG workshops, and is a co-organizer for the 2017 PCG workshop. Adam is also one of the designers of the Indiecade audience-choice award winning game, Bad News.

[SM16] Adam Summerville and Michael Mateas. Super Mario as a string: Platformer level gener- ation via LSTMs. Proceedings of 1st International Joint Conference of DiGRA and FDG, 2016.

[SO16] Sam Snodgrass and Santiago Ontanon. Learning to generate video game maps using Markov models. IEEE Transactions on Computational Intelligence and AI in Games, 2016.

[SSMO] Adam James Summerville, Sam Snodgrass, Michael Mateas, and Ontañón. The VGLC: The video game level corpus.