Intro #
So let me talk to you about my first bit of work, the one that formed the basis for most of my PhD: Automated Game Balancing in Ms PacMan and StarCraft Using Evolutionary Algorithms from Evo 2017.
It even got a nomination for best EvoApplications paper. Somehow. I blame the amazingness of my supervisor.
How and why #
The entire premise of this work was the idea that one could take ALL the numbers that make a game (how much health a soldier has, how fast you can reload your pistol, how high you can jump, how much torque the front left wheel of a Ford Mustang can exert on wet gravel, etc.), pick a few of them and then change them through evolution to get new versions of your game!
…but to what end? To balance the game, of course!
What does balancing the game mean? It means whatever you want it to mean! Game design is an art and what balanced means is up to the artist. My research simply created a tool that let them get from A to B(alanced), as long as they can change the numbers in their game and they have a way of checking that a new version with changes is better than the one they started with.
Games applied to #
For Ms PacMan I put a decent enough AI agent to play the game, then checked how many points they got before losing the game. If I wanted to balance the game to make it easier, I could have my GA change the speed of each ghost, how long a power pill lasted and a couple other numbers, then see if the AI agent would get more points. I also wanted to make the fewest changes possible and I also had some penalties if the game got too easy. Because, you realise, setting the ghosts’ speed to 0 makes it very easy.
For StarCraft it was very similar. Change the training time of some units, how much damage they did, how much damage they could take. Then throw a very good custom AI bot in the arena against the stock bots and see how often the AI bot won. The status quo was “custom bot wins 100% of the time”. We then tried a few scenarios:
- smallest changes to the game to get that win-rate to 0%
- smallest changes to the game to get that win-rate close to 50%
It was good fun and these games proved to be great test-beds for future findings, some completely unrelated to game balance. But those are topics for future posts.
