MrChaos wrote:QUOTE (MrChaos @ May 12 2007, 04:47 AM) 1) Rather then filtering out game based events you should endeavor to be as inclusive as possible i.e. collect it all and let the miners sort it out. Like the famous case of banana prices driving grocery store selection ( whether the individual wanted them or not ).
2) A simple Baysian approach shows great promise in predicting win/loses for this game.
3) Im VERY comfortable hearing in DETAIL what your approach is going to be in tackling this task.
1. The problem with tracking everything is simple: there's too much stuff. Even with a relatively simple game like Quake 3, if you track everything you end up with hundreds of thousands of rows entered every minute, none of it usefully related. This is true of every real-time game.
For example, if I tried to track the positions, thrust vectors, and orientation of each ship every time the physics engine fires... that's a lot of data! (Especially because I have to record the data of everything that moves, not just players!) We could collect it, if we thought it was worth it, but it usually isn't. So, say we only record shot vectors. But
what do we record on firing? Just the shot vectors? Do we want to keep track of live ammo remaining? Ammo left in cargo? Health? Proximity of enemy? Health and ammo of enemy? Too much data!
Instead, you have to pick and choose what data are most interesting.
2. As I mentioned, our interest is less in who will win, and more in how the team interacts. I'm not trying to create an alternate scoring system or anything of the sort. Integrating our data into the game system is
not likely to happen in the foreseeable future.
3. I don't know enough about the game to give specifics as to what we want to track, but here are some examples of things that we might want to track:
What ships players fly. What
roles they fly in those ships.
Actions taken by injured players.
General proximity to other players, ally and enemy.
Team performance after a significant loss or gain.
Overall tactics pursued.
Effectiveness of different weapon loadouts and techniques against specific enemy tactics.
Responsiveness to commands.
Tradeoff of response to command (IE, trying to figure out a bad command versus a good one).
Effectiveness of automated units (turrets, bases, miners, probes, etc) based on some kind of relative position to critical elements...
The problem is, these are just guesses. I won't
know what to track until I've tried it. Learn by experience.
Although if you have specific things that you want to track, knowing about them makes my job easier.
Raveen: I'll definitely put those in. I'll also save max HP of nanned target (and a few other details), so we might be able to calculate context. Context is important.
-Craig
