The idea is based on an existing game: Haxball
It's basically an extremely simple form of soccer. Your bot controls a team of 11 players (circles that move very slowly, but can kick the ball when they touch it), and the first bot to score 3 goals or something wins.
The reason I think it's so cool is that it's physics-based, with continuous states. That's very different from the discrete-state games we've had in the past 3 contests, so your bot logic will also have to be completely different: you can't just minimax your way to victory.
Now there are a few potential problems I can think of:
- 1. I have no idea what the "skill cap" of this game is. It could be that developing an optimal bot is relatively easy, so the top 50 would all be equally good, and games between them would be decided mostly by luck (or be draws).
2. Because the game is so simple, the matches could be a bit boring and repetitive. Maybe adding things like a sprinting mechanic or multiple balls could fix that, but then the game would lose some of its elegance.
3. Since it would use floating point numbers, we would have to deal with rounding errors. Even though the physics are really simple (just circles and straight lines), the game specification may still be pretty complex.
4. Unlike our past contests, this could be a game where beginners have no idea where to begin and this could cause frustration.
5. It may not be that much fun to program a bot for this game. Instead of incrementally making your bot better by adding hacks and heuristics left and right, you would probably develop one big machine-learning algorithm and then tweak its parameters (like learning rate), with long training cycles each time you change a parameter.
6. Robots can't play soccer: http://www.youtube.com/watch?v=7T2aL9XIFQk