Games that "Learn"

[Gilzu]

A friend told me of a really great website ( http://www.pandora.com/ ).

Its basic idea, is that by playing songs and asking you if you like them or not, it chooses the next song. Then again and again until it perfectly fits your taste.

I heard of a few games that "Learn" about the player and what he likes to do - mostly fighting games where the computer learns what kind of comboes or hits the player likes. The same about RTS's where the opponent evolves according to the player.

I thought about how this idea can be applied in games. Check the website out: in each iteration, you can find out why each song was selected. It told me I like "electric rock instrumentation, mild rythmic syncopation, extensive vamping, a vocal-centric and a dynamic male vocalists". Same can be applied to what kind of bonuses the player take most/less, what kind of weapon(s)/move(s) he uses most/less, What kind of levels he likes most ect. ect.

I also thought it might be applied to testing and balancing. Looking at what kind of game elements and times the player used can give you some nice idea what kind of areas you should develop/enhance/fix. Might even be better than asking for feedback - results and records will tell you themselves.

Do any of you know games that "Learn" the player and act accordingly? I know there are some great examples of how you can incorporate this ability in games and maybe try and write one.

Also, if you have some original games/examples of how I can use this idea in a game, I'd be grateful.

[cliffski]

The problem is getting that data back from the player. Most users are pretty shy about sending data from their PC to your webserver, so you need to be sure that they are ok with it. I think it would be a great way to track what features of your game are being used, so you know where future updates should be aimed. Its only my total ignorance of web coding in C++ that stopped me putting this into Democracy.

[MadSage]

One of the first games I remember which learned from the player was Halflife. It's AI was much more advanced than any other FPS game at the time. The AI players learned to outsmart you, and did other cool things like running to get help, then outnumbering you :)

[Fabio]

I thought about how this idea can be applied in games. Check the website out: in each iteration, you can find out why each song was selected. It told me I like "electric rock instrumentation, mild rythmic syncopation, extensive vamping, a vocal-centric and a dynamic male vocalists". Same can be applied to what kind of bonuses the player take most/less, what kind of weapon(s)/move(s) he uses most/less, What kind of levels he likes most ect. ect.At least in Italy laws are pedantic: the player could sue the computer software for having violated his privacy. :D

[lakibuk]

The AI players learned to outsmart you, and did other cool things like running to get help, then outnumbering you :)It's questionable if the AI really learned. Learning means generating new behavior that was not in the system before.

[ManuelFLara]

At least in Italy laws are pedantic: the player could sue the computer software for having violated his privacy. :D
I don't think Gil means sending this data to you, but just using it in real time to change the level layout, enemy behavior, item spawning frequency, etc. in order to fit better that user's tastes. Am I wrong? I think that's a good idea and very achievable if your game is not very complex.

[MadSage]

It's questionable if the AI really learned. Learning means generating new behavior that was not in the system before.

Well I recall reading an article from the developers back then where they stated that it learned, and it did give that impression. I also seem to recall that the game was different when I played, unlike Quake which was the same every time.

[Gilzu]

At least in Italy laws are pedantic: the player could sue the computer software for having violated his privacy. :D

Two things i forgot to mention:
1. When the game learns the player, its main goal is to personalize the game experience rather than share it and create a mainstream taste.
2. I only thought about getting data back in development stages with beta testers. It would be obvius that without asking for proper premission the game will quickly be marked as a spyware. We had a few discussions about getting back data from games a while ago - if you're interested, you're welcome to read these threads, I remember they were pretty interesting.

[cliffski]

It would be good to aggregate data from all players if they would trust you. A game like LUX could theoretically tune its AI strategies by copying the techniques of the winning players, not just on a case by case basis, but globaly. It would mean that the more people play the game, the cleverer the AI seems to get.
I'm not saying its a good or bad idea, but its certainly interesting.

[Fabio]

Dear Manuel, Gil: I really meant it as a joke, I understood the original intention wasn't to send data back to the developer (at least for the release versions).
In Italy privacy seems (and rightly so) a matter of extreme importance, the slightest violation can cause serious trouble (at least theoretically). Talking about AI, I imagined the program getting an obsessive interest about the player's habits, and I drew the parallel. Can you imagine obsessively spying on your neighbour's habits? It would annoy him to say the least. It's funny that something that, if done by a human, would have very serious consequences, is instead very interesting and deserving when made by a software.

Anyway, I see a big potential in your idea.. but I think that, like all AI till today, it is best done as a specific implementation than a generical one. AI has achieved excellent results, in specific fields, although it's still a bit poor as a general tool.

If you allow me to do it, I'd suggest you to read this excellent book on AI: "Artificial Intelligence, A Modern Approach" 2nd edition, by Rossel and Norvig, published by Prentice Hall.

[electronicStar]

One of the identified pleasures of gaming is for the player to learn the game.Not the other way around. If the game adapts to the player it might be difficult to master it.
But it depends on the type of game really.

[impossible]

Well I recall reading an article from the developers back then where they stated that it learned, and it did give that impression. I also seem to recall that the game was different when I played, unlike Quake which was the same every time.
Half-Life having a learning AI is new to me. All of the AI sourcecode is available in the HL and HL2 SDKs, so this could be validated. A quick google of "half-life learning AI" brought up some mention of a engineer in TF2 (which still hasn't come out...) learning about optimal sentry gun placement, but it seemed pretty simple. No trained neural net or anything.

There are some fighting games that having learning opponents (VF4 does I think), but learning is something that works a lot better in a one-on-one game, and you specifically have to play in "learning mode". There are the creatures in Black and White. The problem with a learning AI is that you can't guarantee the player will get a good experience. The AI could learn some clever tricks and end up very challenging (perhaps too challenging) or the AI could end up learning a lot of very wrong things. You also get the situation where players start metagaming the learning system, playing in weird ways to teach the AI bad habits.

[Gilzu]

If the game adapts to the player it might be difficult to master it.

exactly what i'm looking for. Usually, after a week or two that the player achieve mastery in the game, he start to become bored of it. The idea is to keep the challange to the player.

Thats one of the few reasons why people like to play against real life people rather than NPC's.

[Steve Ince]

exactly what i'm looking for. Usually, after a week or two that the player achieve mastery in the game, he start to become bored of it. The idea is to keep the challange to the player.

Thats one of the few reasons why people like to play against real life people rather than NPC's.
That's one of the reasons I don't like playing against other people. :)

[Ronkes]

One of the identified pleasures of gaming is for the player to learn the game.Not the other way around. If the game adapts to the player it might be difficult to master it.

You don't necessarily have to use learning to make the game harder, you can also use it to make the game more balanced towards the player's preferences. For example, when I play a game like Civilization, I like to explore the world and build my cities, but I don't care much for conquest. If Civilization learned this preference of mine, it could adjust my experience by not having computer opponents attack me all the time.

And of course it's also possible to learn the player's habits in order to make the game easier. If the player is very bad at, say, protecting himself against attacks from the rear, the computer opponents could stop attacking him from the rear all the time to make the game less frustrating.

It would even better if the game tells the player: 'Hey, watch your back!' Watch how the player plays your game and then teach him how to improve. That's something I always missed in a racing game. Should I take that corner differently? Is my car setup okay? Am I too aggressive when trying to overtake? Teach me; make me kick ass.

[Chris Evans]

The problem is getting that data back from the player. Most users are pretty shy about sending data from their PC to your webserver, so you need to be sure that they are ok with it. I think it would be a great way to track what features of your game are being used, so you know where future updates should be aimed. Its only my total ignorance of web coding in C++ that stopped me putting this into Democracy.

This has been the great thing about Pow Pow's Mini-Golf (http://www.outsidetheboxsoftware.com/shockwave/webgame.php?game=MiniGolf) so far. Since the game is online and I can pull pretty much any game stat I want. It's interesting to see what modes people gravitate towards and what options people prefer. It's definitely going to help me know which areas to concentrate on when I make updates and improvements.

[Anthony Flack]

I have a long-standing idea for a game consisting of an infinite labyrinth of mini-games. As the player completes the games and chooses their exit, they will make choices that influence whether the next minigame will be harder, easier, more cerebral, less celebral etc.

Somehow.

It's all kind of nebulous still. Maybe one day.

[Illusion Games]

I also have one of those long standing ideas....mine is sort of a cross between an AI that not only learns but is random too.

Briefly - the idea is that you land on a planet that consists of islands that are surrounded by acid. There are valuable things to be mined on this planet. Your initial mission is to protect the mining operation. After a while, you are assigned to explore enemy territory.

This is where the random part comes in. Each game has a different enemy. No comparing notes or reading reviews to discover how this game plays. The player has to explore and use the resources at hand to survive and overcome this unique AI enemy. In that aspect, it is a learning AI because from the AI enemy perspective - the player is just another unit in the game.

Well, that is a rough quick idea. I have to get my other projects finished first!

[djdolber]

One thing that might be somewhat related. I always hated in civ3 that if you gave a generous donation to another civilization, they would always backstab you anyways.:( You could never really have a long lasting friendship between two civs, it would be cool if you could become pals and help each-other out. Now its just a backstabbing game...

But if the game remembered what you did in earlier turns, it might be more interesting... They put some energy into making the civs remember any agreements youve broken, but not the good deeds youve done. I guess thats more related to games that remember and about rewarding different strategys/approaches.