Give me back that beautiful background!
X-pilot is a client-server, multiuser, dynamic game.
In this precise moment
a bunch of humans are wasting their lives chasing each other, playing
Here's a plan for AI-attacking them:
- 1. Put up a million stupid programs playing against each other
to evolve a better one.
- 2. Launch the bastard into the internet and:
- (a) Pass the x-pilot Turing test (fool other players by
- (b) Kill them all!
So, this is too difficult. The game is far too complicated for this
kind of approach, there are too many rules and variations and weapons,
- Simplified, no maze, no gravity, no shields, empty-space plain
- "Second degree" pursuer-evader game (control acceleration instead
of direction). This could be sold to the
(Pacman) style, hack high level behaviors ("attack", "explore",
"run", etc) and inputs ("distance to closest enemy/wall", "being fired
upon", etc.) and adapt a learning algorithm to work on these: Pablo's
- The differential game was solved
- The solution was a little puzzling
- Inputs were hacked to be used by the bot:
- (x,y) coordinates
- Aiming angle
- Angle to nearest enemy
- Distance to nearest enemy
- Density of enemies in neighborhood
- Coordinates of nearest enemy
- Density of bullets in neighborhood
- Behaviors were hacked as well:
- Shoot at nearest enemy
- Approach nearest enemy
- Approach & Shoot
- The differential game studied was of little use when it came to program
- Memorybase learning got one conclusion: SHOOT ALL THE TIME!!
- Future research?
- Find a heuristic mathematical strategy useful in the real
- Is memorybase a valid learning model?
- Improve inputs & behaviors
HERE'S THE PAPER
Comments? Want to play?