|
|||||||
| Register | FAQ | Members List | Calendar | Search | Today's Posts | Mark Forums Read |
| Earlier versions of OOTP: Suggestions and Feature Wish List Let us know what you would like to see in future versions of OOTP! OOTPBM 2006 is in development, and there is still time left to get your suggestions into the game. |
![]() |
|
|
LinkBack | Thread Tools | Search this Thread | Display Modes |
|
|
#1 (permalink) |
|
Moderator*
Join Date: Nov 2002
Location: Brooklyn, NY
Posts: 9,290
|
Building an intelligent GM AI
As I was watching the AI make some dumb moves earlier, I was thinking it would be cool if it could actually learn how to get better. Although the idea expands beyond this, I was initially thinking about the balance between player ratings and stats.
As we know, the option to weigh stats in AI opinions of players went away with the new version (if it ever truly existed). Hopefully this will come back. But what would be very cool is this scenario (note: I'm including "stats" and "ratings" for simplicity's sake; really, there should be things like "stats from this year", "stats from last year", etc): Each AI-controlled team has a relatively random distribution of percentages regarding its consideration of stats and ratings (i.e., one team might base its opinion of a player 50% on stats, 50% on ratings, another 75/25, another 10/90, etc.). So each team values players a bit differently. But here's the good part: At the end of the year, all the distributions shift slightly in the direction of the team that wins it all. So, let's say a team has a 50/50 split, and the team that wins it all is 40/60. Well, maybe the 50/50 team would become 49/51 instead. In this way, the AI would actually be an AI. It would learn. Over many years of play, the various AIs would eventually get more competitive as they all begin to look at players similarly. Eventually the AI would find that "sweet spot" that generates the most success, but there would always be some variation around it due to randomness and other factors. Random fluctuations should happen over time as well so that it never becomes completely immobile (this would be very important, actually). Basically, imitate the way things happen in nature: evolution. And then I started thinking that this same thing could apply to various other strategies in the game, like valuation of stuff/control/movement or other ratings. It might even help create trends over time (e.g. more teams begin to steal a lot when a speedy team keeps winning, or power hitters become more valued when a team of sluggers wins). The game would have interesting variability. This is how AIs are supposed to work: they learn over time and begin to repeat what works and reject what doesn't. I'm probably thinking of stuff that won't be realistic in a game like this for another 10 years, but this is a Wish List item for me.
__________________
THE VERY US ARTISTS - A project for musicians and visual artists My music Currently reading: Thirteen by Richard Morgan "When the trees blow back and forth, that's what makes the wind." - Steven Wright Fjord emena pancreas thorax fornicate marmalade morpheme proteolysis smaxa cabana offal srue vitriol grope hallelujah lentils |
|
|
|
|
|
#2 (permalink) |
|
All Star Reserve
Join Date: Mar 2006
Location: The People's Republic of Yankeestan
Posts: 627
|
I long for the future day when we'll have advanced self-learning AI in video games. What I don't long for is the 4 minutes or so it'll take for AI to become smarter than me.
|
|
|
|
![]() |
| Thread Tools | Search this Thread |
| Display Modes | |
|
|