Red Dead Redemption, GTA IV, Limbo. Works of creative genius by some of the most imaginative, most talented video game designers on the entire planet. All made by humans, of course. One PhD student at Imperial College, London, however, is on a mission to show that in the future some of our video games might be made, not be people, but by computers.
Step forward ANGELINA, one of the early AI pioneers of this new realm.
Now, before we get too excited, ANGELINA isn’t a true sentient AI encased in a metallocerramic carapace who goes out into the real world and gets inspired by Mozart symphonies, glorious sunsets, and the street art made by vomiting drunks, then goes back to her robotic garret and codes a masterpiece. Oh no. ANGELINA is merely a computer program who knows very little, if anything, about her digital creations.
So, how can she create games?
Answer: by leveraging the powerful force that is evolutionary computing. Let’s unpack that a little. First, what is evolution? Well, in the natural world, evolution is the chemico-biological process by which species evolve and diversify. Anything that produces offspring is in thrall to the power of evolution. This happens because when a biological entity reproduces, either sexually or asexually, the offspring are usually slightly different from their parents. Those which have developed traits that make them better at surviving their environment will be more likely to produce more offspring themselves, thus propagating their succesful traits. Those whose traits are a hinderance rather than a help will be less likely to reproduce and the trait will not be passed on.
On a human timescale we don’t see the effects of evolution in most organisms because it takes many generations for these traits to propagate. However, by mimicking evolution inside a digital space where we can speed up the process by many orders of magnitude, we can evolve, not just digital organisms, but anything that can be represented by a “genetic” code and tested in whatever way we see fit: aircraft fan blades, Prisoner’s Dilemma strategies, and, you guessed it, even video games!
While in nature the “fitness function” is encapsulated in an organisms ability to survive and reproduce before dying, the fitness function for population members like fan blades and video games is less clear, and is one of the more difficult parts of creating an evolutionary algorithm. For example, what makes any given fan blade “good”–and therefore a stronger candidate for reproduction where its design can be finessed–will be a subjective blend of how much material is used, the length of the blade, its aesthetic value, perhaps, how much turbulence it generates etc etc.
In a domain such as video games this assessment might be even harder; one person’s Call of Duty: Modern Warfare 2 nirvana might be another’s Shaq Fu horror.
ANGELINA creates games by breaking down a video game into constituent populations that can undergo this evolutionary process, with each game of every generation tested by a simulated human player that makes a numerial assessment of its “fitness”. For example, level-layout, enemy behaviour, and power-ups are each seperately evolved before being combined. The games that score highly will then have their constituent parts “reproduce” which will result in a new generation of slightly different levels etc, some worse than before some better. Those that are better will form the basis of the next generation, and in this way ANGELINA optimizes her collection of games, eventually–hopefully!–producing something that a real person can play and enjoy.
Judging by ANGELINA’S creations I don’t think Peter Molyneux–creator of Powermonger, Syndicate, and Magic Carpet, among many other classics–has to worry, but maybe some of the Indie studios churning out dime-a-dozen game apps for Android and co. should!