For my final project for Nature of Code: Intelligence and Learning, I would like to investigate how generic algorithms can be used to design the landscape in video games. (Despite the fascinating prospect of convolutional neural network and mesmerizing ways of training and using pre-trained CNN models, I’m actually quite interested in the first half of this course, where we were designing the evolutionary processes and setting the parameters for “natural selection”, and visualize the evolution over time.)
On one hand, because game design concerns with chances and rules that affect how players behave in the game and how difficult it is to accomplish the task or to win the game, I think it would be very interesting and rewarding to try to translate those considerations and design decisions of video games into machine learning algorithms.
On the other hand, I came across Lev Manovich’s Navigable Space (1998) while in undergrad, and was very inspired. He compared and contrasted Doom and Myst (both published in 1993), and how the idea of “space” is different: in Doom, the player is prompt to navigate the space and overcome enemies in the shortest time possible, and there was even an “open source” feature whereby users can create their own advanced levels; in Myst, the space is predefined, and the player “wanders” in order to find necessary information to decode the challenges. But the two games are in common in terms of how the player’s experience is based on “spacial journeys”
While the blueprints of video game space is defined by game designers, based on their skills, experiences and user tests, I wonder if genetic algorithms can help create a controlled blend of geographical complexity, number of challenges(enemies), rewards, and over difficulty in the design of a “navigable space”. What’s more, algorithms can also be used to “test” the generated game space, allowing rapid iterations.
This project proposal is also inspired by the idea that gamenification, or thinking about real-world phenomenon by abstracting them as games, has been an fruitful way for people to better understand things happening around them. For example, Guy Debord famously designed the Game of War at the late stage of his career (1987), and recreated the militant situation between north and south France in a particular history period. As media scholar Alexander Galloway describes it:
“He(Debord) positions chess firmly in the classical period of kings and corporal fiat, while the Game of War belongs to a time of systems, logistical routes, and lines of communication. In chess, spatial relationship between pieces are paramount; the ‘knight’s tour,’ for instance, serves as a classic mental projection of pattern and recombination. In the Game of War, Debord maintained this attention to spatial relationships, but added a degree of complexity. The ‘liaisons’ in the Game of War are not simply the projections of possible troop maneuvers, but a communicative apparatus linking together far-flung fighting divisions. If chess’s king is an intensive node, one that must be fortified through the protection of its allied footmen, then Debord’s arsenals are extensive nodes. Yes, they too must be protected, but they also serve as the origin point for a radiating fabric of transition. In chess ‘the king can never remain in check,’ but in the Game of War ‘liaisons must always be maintained.'”
In fact, Galloway himself had recreated Debord’s game in the form of an online two-player game, written in Java, and won a Golden Nica from Ars Electronica…
Due to the complexity of the subject, I think it would be better to first star with some simpler games, such as recreating the PEC-MAN maze using algorithms.