Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Jay]

Cleverbot is just a glorified, online, MegaHAL.  Trying to make it sound even slightly intelligent is...  a lofty goal, to say the least.

[Kay12]

Well, it's a toy AI. I don't personally consider building a perfect chatbot to be worth the effort - not that it wouldn't be interesting as an academic curiosity, but... other kinds of AI have actual practical uses.

Now that we have an AI thread, by the way... news from a few weeks back: RTFM applies to intelligent agents too, it seems.

[counting]

IIRC, I heard a lecture about gaming AI. It's goal is not play to win, but play to lose. It has to make players (great) winning chances, and not exploit extreme strategy. (letting the players to exploit them). You have to limited the search depth in order for an AI to function quickly. Most of the CPU times are not using to process gaming AI as well. Calculation goes to determine the interactions of objects (physics engine).

A link to Google Tech Talk about Civ's AI.
http://www.youtube.com/watch?v=IJcuQQ1eWWI

[alway]

Yep. Ideally, game AI should lose, but in such a way as to give the player a challenge and in such a way as to reduce computation time required.

[counting]

Learning is just one side of the AI (regardless a very important part and heavily researched area). The other half is how it behaves. You probably all have the feeling when playing games, that you almost want to shout at the AI, since it's so predictable. When you giving a similar conditions for it, it will always behave that way (or choose several methods but in very limited paths, with predictable outcome probabilities).

One thing not so well known in NN research, is that it can behave "creatively". I've building a NN model before using a RNN structure for it to "perform" short music pieces (up to 255 notes), by giving it enough pattens. It can generate some interesting and unseen(unheard) outputs. Very hard to predict a NN based AI, as if it has a will of its own. On the other hands, most gaming AI are just rule based agents, not even combining the ability to adapt. (Some are). Not even GA is widely used in gaming AI. (Even if it's so simple to implement.)

[Kay12]

IIRC, I heard a lecture about gaming AI. It's goal is not play to win, but play to lose. It has to make players (great) winning chances, and not exploit extreme strategy. (letting the players to exploit them). You have to limited the search depth in order for an AI to function quickly. Most of the CPU times are not using to process gaming AI as well. Calculation goes to determine the interactions of objects (physics engine).

A link to Google Tech Talk about Civ's AI.
http://www.youtube.com/watch?v=IJcuQQ1eWWI

You've got a good point. Also, it all depends a bit on the purpose of the AI. For Civ IV, the creators seemed to want a quite realistic simulation of an actual government - they seem to follow some ethics that humans, ironically, don't tend to practice in games like this: only a few leaders will declare war on their friends, defensive pacts are always cool with friends even if their military sucks, liberated colonies always love their master, you don't attack someone simply because they're winning etc...

For Civ V, they chose something like a middle point between "simulation" and "game" AI. Civ V AI tends to hate players who do better than them, gets angry if someone builds too many cities and even "abuses" the research agreement costs to drain money from the player they're about to declare war on. Oh yeah, and it plays to win.

[DrPoo]

AI to win > AI to loose, games are simply not fun.. its like first time i played warhammer and the other guy was deliberately trying to loose

[Kay12]

You got that a bit wrong. For example, it wouldn't be very hard to create a chess AI that plays well enough to be practically unbeatable. But what's the fun in playing against an opponent that plays perfectly? A typical recreational chess AI should have some flaws that make it beatable by good playing.

Also, optimal play isn't always even the developer goal. For example, Civilization AIs feel more "human" when they drag in some "emotions" - reluctance to give up cities (even ones that are almost worthless), not making defensive pacts with players they don't like even if they're strong, declaring wars when they have almost no chance to win...

Another example would be the Texas hold'em minigame from Red Dead Redemption. You're supposed to be playing Poker with a few angry gamblers, not computers that can instantly calculate their probability of winning. If you want to simulate humans playing Poker, you could also add stuff that's downright useless for actual winning - trying to take revenge on another player, for example.

[alway]

AI to win > AI to loose, games are simply not fun.. its like first time i played warhammer and the other guy was deliberately trying to loose

In that case, though, the AI was not flawed because it was meant to make you win, but rather because it failed the second point: giving you a challenge in winning. Depending on the AI, the requirement to use as little resources as possible can limit them quite a bit. You see this in AI who cheat; either getting free stuff or increased money/decreased build times, things like that. Those AIs are essentially really crappy and need to cheat simply to pose the player a challenge.

[counting]

Meta knowledge and meta-meta-knowledge, or anything higher, is something makes A.I differs from a simple Q/A reflex chatbot. The more freedoms in an AI's meta-structure, the more "human like" it will be. Like in Civ's AI, it has internal state of hostile, neutral, friendly, they are meta-data. And in recent chatbot, they all tag words with emotions. So perhaps a better AI is NOT the one who appears "intelligence" and having "smart-ass" conversations, but a self-talking like a stream of consciousness with complex meta-data structure flows of internal states.