| Author |
Message |
Richard Hutnik (Richardhutnik)
| | Posted on Monday, December 16, 2002 - 6:12 pm: | |
Howdy Jeff and crew...
I recall reading that you ended up tweaking certain values in Zillions to play certain games better. I have a question regarding this: Would it be possible to have it so that an AI at Zillions could end up, through repeated play, end up autotuning itself to get better? I am sure some games value movement more than others.
Thanks...
- Richard Hutnik |
Jeff Mallett (Jeffm)
| | Posted on Saturday, December 28, 2002 - 8:14 pm: | |
Yes, it would be possible to do something like this (although it's not in the current version.) We've done similar things for our chess programs. The way it usually works is that you vary a single parameter and then play that program against the original version a statistically significant number of games, say 100. If the new version wins most of the time, then you keep the change. Then you vary another parameter and repeat.
There are negatives to this approach. One obvious one is the large number of games you have to play. A not so obvious one is that you may only be hitting a local maximum and have no guarantee of ever achieving optimal values. For example, there may be an interplay between a number parameters and only by decreasing one while simultaneously increasing a couple others will you get a better result. |
Keith Carter (Keithc)
| | Posted on Friday, January 03, 2003 - 11:53 pm: | |
There are some limited possibilities for pitting one AI evaluation within a Zillions game file against another AI evaluation. I am working on a game with Dan Troyka that has more than one way to script the win condition. By giving the two sides different but equivalent win conditions I can set Zillions to play against itself and get a measure of which win condition leads to stronger play by Zillions.
Another way is piece evaluation comparisons. A piece's value can be increased by creating a move type and giving it moves descriptions that are not available in the turn order. If done for a piece type for only one side then Zillions will evaluate and play the two sides differently. This is only helpful for improving AI if Zillions is undervaluing a piece type. |
|