|
|
Post by nawglan on Feb 23, 2009 10:43:34 GMT -5
Heh. Noticed it's been 3+ years since I created my warbot player on warlocks. The state of the AI has not changed to my knowledge, so I thought I'd bring this crazy idea to the forefront once again.
This time the topic is inputs. What factors into your decisions about what gestures / targets you enter at any given time.
List I have thus far:
health of yourself and opponent
gesture history of yourself and opponent
affected by list of yourself and opponent (charmed, etc.)
monster list (type, health, owner, target)
I'm assuming that from the gesture history, a list of possible spells can be derived. Though, I was not thinking of using that list as a direct input.
Have I missed any inputs?
|
|
|
|
Post by maknud on Feb 23, 2009 11:24:50 GMT -5
Targeting history of you and your opponent. Vital vs the paraswarm.
|
|
|
|
Post by mikeEB on Feb 23, 2009 14:21:13 GMT -5
Inferred gestures from spells cast during invisible/blind turns.
Also, I take the opponent's ELO into account, but I wouldn't expect the bot to.
|
|
|
|
Post by maknud on Feb 26, 2009 22:12:07 GMT -5
Why wouldn't the bot? Presumably it's trying to maximize something; likely that something is expected ELO change for the game. Knowing whether a DD gives points or takes them away is useful.
More subtle but not too hard to do roughly is e.g. valuing health trades more when against higher-ELO foes, since you don't want their superior knowledge of position to have as much chance to come into play.
|
|
|
|
Post by nawglan on Mar 2, 2009 13:20:58 GMT -5
I'd like to play a couple games for ai training purposes. The games would be open handed (meaning that all gestures are declared in chat as well as targets if the target is not default) and would be VF, but should be played for blood. No slacking. I'll be using these games to create training data, and they will be archived so that anyone else can use them too.
No rush on these. Only need to play a few games, but having a bunch won't hurt. If you are interested, send me a game invite. If you don't want to play me, but want to help... be sure to declare the gestures / targets and send me the game number.
|
|
|
|
Post by xade on Mar 2, 2009 17:30:59 GMT -5
always happy to help! Too lazy to send invites though...  |
|
|
|
Post by mikeEB on Mar 2, 2009 17:37:59 GMT -5
Why wouldn't the bot? Presumably it's trying to maximize something; likely that something is expected ELO change for the game. Knowing whether a DD gives points or takes them away is useful. More subtle but not too hard to do roughly is e.g. valuing health trades more when against higher-ELO foes, since you don't want their superior knowledge of position to have as much chance to come into play. Actually, I mainly use ELO as a predictor of how likely they are to botch a para-FoD, permanency or poison defense. For example, if I end up with DFF/XXP, I will go for bolt more often vs. high-ELO players and para-FoD more often vs. low-ELO players. |
|
|
|
Post by nawglan on Mar 2, 2009 18:58:10 GMT -5
Meaning, you'll more than likely Para-FoD me. 8)
After a second look. Only need gestures to be declared if you are under a mindspell or invis / timestop. Rest of time, I can grab them from the summary line.
|
|
|
|
Post by ourjake on Mar 3, 2009 18:55:40 GMT -5
you can invite me or syd, we will help with the stuff and are usually fairly quick as far as submitting moves so it can be gathered quickly
|
|
|
|
Post by nawglan on Apr 1, 2011 23:54:55 GMT -5
Well. I decided to resurrect my AI dreams... And to my chagrin, my latest attempt came out a paramancer! I started with LH/RH B, B for both players. I took the highest rated pick for what it predicted the opponent to do, and what it would do. (admittedly, it assumed a very stupid opponent).
Game Options: Maladroit, ParaFC
Player1 (the bot) - affected by maladroit turn 4, targets nobody turn 9
LH:BPSDFFDDSPFP
RH:BDPPFFFFFFFF
Opponent - charmed by bot turn 5, bot then para'd lh, bot then switched to rh
LH:BPPPPSFCCCDP
RH:BDSFF-PPCSWP
|
|
|
|
Post by mikeEB on Apr 2, 2011 0:01:37 GMT -5
How is it generating the predictions again?
|
|
|
|
Post by nawglan on Apr 2, 2011 0:18:27 GMT -5
It looks at the last n gestures and has a lookup table that it uses to see if it's seen those gestures before. If not, it reduces n by 1 and looks again. and so forth until it finds a hit.
|
|
|
|
Post by salvor on Apr 10, 2011 8:28:30 GMT -5
I would like to play training game if you'll invite me(lack of slots).
|
|
