Can anybody here give a reference to commercial AAA games that implement a machine learning AI?
\$\begingroup\$
\$\endgroup\$
3
-
\$\begingroup\$ Like the current answers say, I doubt that you'll find any examples. The closest you might find are titles with a sliding scale of difficulty based on player actions and success, but the actual AI is pre-built. \$\endgroup\$– Patrick HughesCommented Sep 19, 2011 at 17:59
-
2\$\begingroup\$ Not a useful question, what does it matter? \$\endgroup\$– HouseCommented Apr 13, 2012 at 22:55
-
\$\begingroup\$ This question appears to be off-topic because it is about compiling a list of games. \$\endgroup\$– user1430Commented Jan 23, 2014 at 15:44
Add a comment
|
6 Answers
\$\begingroup\$
\$\endgroup\$
1
The most common approach is to train the AI off-line or pre-release, and to ship that final result. As such the game arguably doesn't contain a machine-learning AI.
This is because machine learning can traditionally take many hours of training to produce a usable result, and as it's by definition emergent, it's probably a good idea to put any result through a full QA pass to make sure there aren't any exploitable edge cases.
For the above reasons, and because gameplay changes and balancing even late on might mean that the AI has to be retrained and retested, not many studios use this approach.
-
3\$\begingroup\$ This answer makes sense, but in what world does regular AI not have edge cases :) \$\endgroup\$ Commented Sep 19, 2011 at 22:40
\$\begingroup\$
\$\endgroup\$
Forza's Drivatars are an example of learning AI in a shipped game. They watch your driving to learn how to copy your style. There's lots of information in that link.
\$\begingroup\$
\$\endgroup\$
1
The only game I knew that I think may use machine-learning is Tekken 5: Dark Resurrection for psp. in that dojo mode game AI could learn how you play with each one of characters and it created a ghost of you playing that specific character. later you could fight with yourself or give your shadow to your friend and let him to fight against your shadow.
-
\$\begingroup\$ Yes 2D/3D fighting games (even old ones) are perfect example of self learning machines : the game analyse your actions and movements and based on this can predict what the next ones will be (most players always do same tricks). The AI will next do the appropriates movements to counter them. \$\endgroup\$– tigrouCommented Jun 2, 2012 at 18:51
\$\begingroup\$
\$\endgroup\$
I don't know if it counts as a AAA title, but Darkwind uses genetic algorithms to train AI drivers. There's also a newer paper about using player behaviour to modify pathfinding in complex terrain, which you can read here.
\$\begingroup\$
\$\endgroup\$
It's not a AAA title, but more of a research game, you create a squad and then train it by manipulating the enviourment, the ai is based on neural net's and the learning is done through GA. NERO
\$\begingroup\$
\$\endgroup\$
Galactic Civilizations 3 will use a networked approach to AI Machine Learning. Player tactics and strategy will be uploaded to the company servers for analysis. From the Gal3 website faq (http://www.galciv3.com/game/faq):
"Q: : Will the game require a network connection to play?
A: No. However, some of the AI learning mechanisms will require access to the Metaverse in order to perform analysis on player strategies. This will be optional and able to be disabled by the player but does require an Internet connection."
No details on if this is real time, delayed or used in game updates, but the GalCiv games have always been noted (in reviews) for the solid AI.
