It seems machine learning can be used for everything these days. Could you use it for PCG as well? Yes indeed. But the challenges posed by game content like maps, levels, characters, and rules pose different challenges than images, music, text, and other types of non-game media that machine learning has so far been used to generate. In particular, we have the challenges of playability and challenge. I will discuss the particular ML challenges for generating game content and give a number of examples of what we call PCGML. Methods that have been used for this include n-grams, Markov Chains, LSTM recurrent neural networks, generative adversarial networks, autoencoders, and neuroevolution. I will also discuss how PCGML can be combined with search-based methods in order to overcome some of the challenges associated with balance and playability, for example using latent variable evolution. Finally, I will discuss the broader range of methods and paradigms in machine learning, and discuss some unexplored opportunities. For example, could we see PCG as a reinforcement learning problem?
In this talk, I will be addressing the questions of how we can construct models of designers and players and how we can use such information to generate appropriate game content for them; this is game content that both satisfies necessary functional properties and incorporates appropriate aesthetic elements. I will be covering methods derived from procedural content generation that can be coupled with player models to yield entirely novel and personalised content for each player or designer. Further, I will be showcasing ways we can debug the experience attributed to each content type we design in a semi-autonomous or designer-assisted way.
Website: http://yannakakis.net/
Procedural content generation is found in every corner of game development – in art, music, design, writing, QA and more besides – but our ways of thinking about procedural generation have stagnated, and a lot of game developers use procedural generation in the same conservative ways. In this talk, we’ll look at how generative systems and randomness affect games in different ways, and what that can teach us about using procedural generation to discover new kinds of game, build new kinds of tool, or do new kinds of AI research. Topics covered include procedural content generation, automated game design, and computational creativity.
Website: http://www.gamesbyangelina.org/
Procedural content generation often relies heavily on human knowledge. The increase in availability and variety of datasets encouraged the use of open data as a source of information and inspiration to automatic content generation. Data games use (open) data to automatically generate game content, leading to interesting juxtapositions of information, gameplay, and visualizations. This talk will cover methods to interact with and transform data into game content, how different data types and content types can affect a final result, and the pitfalls and advantages of using open data into a product.
Website: http://gabbbarros.com/
The specific outcomes of procedural content generation are hard to predict. Obviously this is one of the reasons to use PCG in the first place, but in the context of commercial game development this can become something of a problem. Ideally, one would like to be able to direct PCG and still be surprised by it. In this lecture I will discuss how we deal with this issue in the development of Unexplored 1 and Unexplored 2. These games use grammar based content generation using Ludoscope that mimics a top-down design process not unlike the process followed by human designers. To achieve the desired results requires the right balance between randomness and control, many iterations, and close collaboration between a designer and the tool.
Future Unfolding uses procedural generation on all levels of the game, from decorative placement of plants to layout and ordering of world segments. This talk will go through the methods we used and show some of our internal tools. In particular our level editor, which allows us to combine generated and manually designed areas in the game.
While we often think of content as facets of games that we see, hear, or interact with (e.g., levels, narrative, graphical assets, or music), some games feature player-generated content as a core mechanic. For example, Super Mario Maker is a game where players create and play each others levels. Another game is Hearthstone, where players must either pre-plan decks before games or plan them on the fly, depending on the particular gameplay mode. In both of these cases, a large component of the game requires players crafting different elements to play successfully. Interestingly, deckbuilding is also a significant challenge in AI. In Hearthstone, different gameplay modes and formats restrict the cardsets in a variety of ways (e.g., Standard format, Arena, Tavern Brawl). In the least restrictive format (i.e., Wild), there are over 6.12 × 1020 possible decks. However, most often there are only a few competitive decks that dominate the metagame. Computationally exploring the space is difficult, so players rely on their intuition and the success of others to both build and play these decks. However given the size of the space of possible decks, it is likely that there are many competitive decks that have yet to be discovered. This talk will survey some approaches to deck building in Hearthstone, and discuss some current challenges posed by the game from a procedural content generation perspective. The speaker welcomes any potential collaboration with interested students.
Many tools are built to help professional people turn their visions into reality, but some creative activities are *autotelic*, that is, creativity experienced for its own playful purpose. Can we design systems to help these casual users engage with their creative sides?
More speakers will be added.
