Over the course of my research I have come across a number of testing domains, both official and unknown. This post intends to list them, both as possible testing grounds for my agent, and for others to use.
Blocks World
This one needs no link or introduction. If you have come across this post, Blocks World is where you should probably be starting. It’s easy to grasp, but not necessarily simple to overcome (for an agent).
Ms. PacMan
My first domain. Agent plays Ms. PacMan over a number of levels. Domain isn’t easily grasped as a relational domain, but will soon be.
StarCraft
Or for this matter, any RTS. But StarCraft is currently where my efforts are going, as there is an international contest running for it currently, and it already has tools for interfacing with the game and creating a bot.
Game of Intelligent Design
A collection of tasks for a script-based agent to solve. Currently, I am unsure whether it’s simply a test of creating a smart, one-shot agent that completes the problem or if the agent is capable of storing knowledge over a number of iterations. Some of the tasks seem to require that the agent has some form of knowledge evolution. The problems again aren’t easily put into a relational map, but I’m sure it’s possible. The problem lies with the numbers in the environment.
Car Racing
A car racing competition is also being run at the moment, which would be a fun environment to attack. While I’m not entirely sure how the environment is given, I figure it could be relationalised fairly easily.
ACM Queue Challenge
While the competition is over, I reckon the software should still be available. A simple, yet complex game of trying to capture counters using a ‘line-drawer’ and some ‘pushers’. The rules are simple, but an optimal play strategy is near-impossible. Especially when an opponent is involved.
Google AI Challenge
Again, this competition seems to be over, but it is another possible environment. The problem lies in how to implement it.
Mario AI
The Mario AI competition seemed to start in 2009, or at the very least it caught my eye then with the popular video of the A* Mario solution. The Mario world is represented in Java, making the environment easy to rig up and test with the agent. Perhaps this could be part of my batch? StarCraft, Ms. PacMan, Mario and Tetris? A thesis on computers playing computers.
PyLab
A testing environment developed by a group of RL researchers. It contains many testing environment and other things, including RLGlue support. Not that that’s much help for my relational system, but it did begin with RLGlue, so it shouldn’t be too difficult to write a proxy environment.
CIG Competitions
Interestingly, nearly all of the above domains are competitions held by CIG (Computational Intelligence in Games) conferences.
There are probably heaps of other ones too.