|
Post by dni on Sept 27, 2007 4:02:36 GMT -5
2 nawglan: LOL Your AI is so funny I enter DSF/PSD at opponent's hands, and WWP/SPF at my hands, and what result i get? Unbelievable! P/P!
|
|
taliesin
Ronin Warlock
Grand Master
Posts: 156
|
Post by taliesin on Sept 27, 2007 5:57:13 GMT -5
2 taliesin: So what's about psychic influence? Making a game with one HP? Is you need to attack with one hand? Making shield or do a spellchain? The evaluation method is only half of game key... The psychological part of the game? That has to be handled by randomly selecting one choice or the other. However, the evaluation is still almost all of the work here - once you've got approximate values for the different outcomes, and you can calculate how much predictability will cost you, then you can work out what odds you'll set on jumping each way. The neural net isn't going to solve the predictability problem, you need to compensate for that explicitly.
|
|
|
Post by dni on Sept 27, 2007 9:28:42 GMT -5
2 taliesin: It's much smarter to examine player's habits and play against that player. Random is unpredictable so player can't know what AI will select, but it's probably possible for human player to play on chance calculation so computer will prefer node with more probability of win. Playing against random, huh? Also it's interesting to make AI habits and AI assotiative decisions (so maybe based on neural network, for example). It gives a chance for human player to play not with unlive machine. By the way, all strict algorithms which evaluates 'weights' and makes a decision based on it, will totally lost to smart human who can make a mind trap for AI.
|
|
taliesin
Ronin Warlock
Grand Master
Posts: 156
|
Post by taliesin on Sept 27, 2007 10:51:31 GMT -5
2 taliesin: It's much smarter to examine player's habits and play against that player. Random is unpredictable so player can't know what AI will select, but it's probably possible for human player to play on chance calculation so computer will prefer node with more probability of win. Playing against random, huh? No, it's not, because in almost all cases you will have insufficient data to make informed decisions about the player's habits, and furthermore your AI then becomes vulnerable to players simply varying their play against it. Your AI becomes predictable. This is why many of the top players, when faced with a true 50-50, will toss a coin and use the result - to make sure they don't lose out through being predictable. If you're using minimax, you're always looking at the worst case - in this case, how predictable are you? - and seeking to minimise it. Also it's interesting to make AI habits and AI assotiative decisions (so maybe based on neural network, for example). It gives a chance for human player to play not with unlive machine. Sorry, don't understand what you're saying here. By the way, all strict algorithms which evaluates 'weights' and makes a decision based on it, will totally lost to smart human who can make a mind trap for AI. False; how do you think chess AIs work? And Warlocks is simpler than chess. Humans are now hopelessly weaker than draughts and checkers AIs.
|
|
|
Post by dni on Sept 28, 2007 1:41:11 GMT -5
2 taliesin: It's right, i think it will be interesting to try of make it. Only interests, not neednesses. Player who will varying their play too often, make strong vulnerability for himself. You never expect in this case how AI will modify his strategy based on your moves and game history. Flexible methods never can be predictable. You can't feel the changes in this case. I know that one of variants but not best of all is random. It's may compared to despair when you give your future to the hands of your fate (i don't know, coin is really random or it turns by a hand of god? but it's philosophical offtop ). Try to play a game 'odd/even' with somebody and you feel the sense. (By the way, rules of 'odd/even' is elementary: two players write numbers simultaneously, and if sum of them is odd, one player wins, if sum is even, another player wins). I prefer to use a method when all cases are possible but with different probabilities. For example, AI evaluates that one of move may give him 75% chance to win, and another -- only 25%. IMHO, the decision will be maked not for first variant. It needs to make random value with (0.75, 0.25) distribution or apply some kind of distortion function to probabilities so each case may happens. I'm sorry too, sometimes i have trouble in translating. I mean that it's interesting to make AI which can have own real habits. This habits may shift from time to time. Secondly, it's interesting to make not logical, but associative AI decisions. This can make a chance to human player to play with AI, cause it's not interesting to play with non-live machine which simply calculates its move. Make an intellectual, not a calculative AI -- it's really a great task! Hm-m-m. You feel? I think it maybe softer. The chess it not right comparison. The chess has sequential moves, as there as warlocks has simultaneous moves. Unlike chess, warlocks has strong element of bluff and psychological influence. The right comparison may be some card or board games (like 'baccara' or 'poker' in casino, etc; you know good program which can play with human player in poker?). Quite simpliest game is 'odd/even'. It's simplier on position description in each move but harder on quantity of variants (chess never have 64 choices at move). Plus it needs to take into consideration of simultaneous-move game structure. Sequential games are fully calculative (theoretically, i mean), unlike simultaneous.
|
|
taliesin
Ronin Warlock
Grand Master
Posts: 156
|
Post by taliesin on Sept 28, 2007 6:39:13 GMT -5
Player who will varying their play too often, make strong vulnerability for himself. You never expect in this case how AI will modify his strategy based on your moves and game history. Actually, you'll pick up a feel for it sooner or later. It's sometimes possible to get a human player with a very straightforward switching back and forth of attacks when he can only shield one target, and you're attacking him, then his monster, then him again. Humans recognise after encountering this once or twice that they're vulnerable to being predicted and learn to mix it up. AIs aren't capable of that level of abstraction. Flexible methods never can be predictable. You can't feel the changes in this case. You mean they can never be perfectly predictable. They can, however, be mostly predictable. Humans learn faster than neural nets. They're likely to adapt faster to the neural net than the net does to them. I know that one of variants but not best of all is random. It's may compared to despair when you give your future to the hands of your fate (i don't know, coin is really random or it turns by a hand of god? but it's philosophical offtop ). Minimax, dude. When you write an AI, its goal is to minimise the score achievable against it, not to hope the other player will hand it the best possible outcome. You assume your opponent is better at predicting and work to minimise this. It's the way AIs work. I prefer to use a method when all cases are possible but with different probabilities. For example, AI evaluates that one of move may give him 75% chance to win, and another -- only 25%. IMHO, the decision will be maked not for first variant. It needs to make random value with (0.75, 0.25) distribution or apply some kind of distortion function to probabilities so each case may happens. Yes, I said this already, and in slightly greater depth. I mean that it's interesting to make AI which can have own real habits. This habits may shift from time to time. Secondly, it's interesting to make not logical, but associative AI decisions. Interesting, but not practical if you want a competitive AI. This can make a chance to human player to play with AI, cause it's not interesting to play with non-live machine which simply calculates its move. Make an intellectual, not a calculative AI -- it's really a great task! You are aware that it's also an impossible task, given current technology? (I'm assuming what you mean here is to make a pattern-matching AI instead of a state-space search one). The chess it not right comparison. The chess has sequential moves, as there as warlocks has simultaneous moves. Unlike chess, warlocks has strong element of bluff and psychological influence. Warlocks is very, very much more like chess than poker, and while bluff matters, it matters much much less than the ability to evaluate your position and see the moves that will improve it. The right comparison may be some card or board games (like 'baccara' or 'poker' in casino, etc; you know good program which can play with human player in poker?). Quite simpliest game is 'odd/even'. Warlocks is almost directly comparable to a chess game in which both opponents write down their moves simultaneously, hidden from each other, and then reveal them. It won't be pure chess, but the people who play chess the best are the people who're most likely to do well at this game. It has no relation to card games: you have a current board state that affects the moves you can make, and hence position can be incrementally improved. Poker has no such concept of position. Every hand is played under the same circumstances. Also, Pokibot is rather better than most casual poker players, though it's no match yet for the really good ones. It's simplier on position description in each move but harder on quantity of variants (chess never have 64 choices at move). Let's say almost never, but it's worth pointing out that Warlocks doesn't have 64 choices per move either: there's never a reason to select a null gesture. That brings it down to 49. There's never a reason to select a half clap (assuming you control your own spellflow, and have all the choices available) - that brings it down to 37. I could mention the relative uselessness of stabs, but I'm going to move straight on to disruptions - Amnesia gives you just 1 choice, Fear 4, Maladroit 7 (discounting the null again), Charm 7, Paralysis 7. Chess has no equivalent, and the combinatorial explosion is thus much more potent - chess generally has 20-30 options available each move through most of the early game. Plus it needs to take into consideration of simultaneous-move game structure. This is much less important than you're making out. It doesn't require any structural changes. Most of the work goes into constructing an AI that would be a killer if the game were wholly sequential, with only a little at the end needed to translate the relative evaluations of positions into a randomised selection from a choice of moves rather than coming up with a single one. Sequential games are fully calculative (theoretically, i mean), unlike simultaneous. Yes, but the tactics used for the sequential game can work in a simultaneous game if the concepts of positional advantage still hold true.
|
|
|
Post by freesoul on Sept 28, 2007 7:40:35 GMT -5
"Poker has no such concept of position. Every hand is played under the same circumstances." Position definately exists in tournament style poker. Much depends on who (in style and skill) sits next to, and across from you. This will impact how you play (which changes when people get knocked out). Also, the turn you get to act plays a big part... whether you act first and have a bunch of unknown hands acting after you, or acting last and getting a read on everyone as well as knowing that no one will raise after you. Finally, I get to chime in on all of this AI gibberish
|
|
|
Post by Slartucker on Sept 28, 2007 7:48:18 GMT -5
Once again I will implore anyone who feels like arguing about this topic to READ the AI piece on the Refuge. There's no reason for all of this to be retyped again and again...
|
|
taliesin
Ronin Warlock
Grand Master
Posts: 156
|
Post by taliesin on Sept 28, 2007 8:47:18 GMT -5
Position definately exists in tournament style poker. Much depends on who (in style and skill) sits next to, and across from you. This will impact how you play (which changes when people get knocked out). Also, the turn you get to act plays a big part... whether you act first and have a bunch of unknown hands acting after you, or acting last and getting a read on everyone as well as knowing that no one will raise after you. Okay, perhaps that's badly worded on my part. My point was that chess and Warlocks have complex variations in position that cause all moves to be evaluated as "good" or "bad" solely on the basis of the context in which they appear, to a degree that card games do not share, the positional context changes rapidly, and these positional factors are largely independent of who is playing the game.
|
|
taliesin
Ronin Warlock
Grand Master
Posts: 156
|
Post by taliesin on Sept 28, 2007 8:58:50 GMT -5
Once again I will implore anyone who feels like arguing about this topic to READ the AI piece on the Refuge. There's no reason for all of this to be retyped again and again... We're currently arguing a substantially different topic from the Refuge AI piece, despite the references to chess. dni's argument appears to be: 1) Warlocks is played simultaneously, not sequentially, and this changes everything; we need to look at card game AIs instead of chess AIs for our model. 2) It is feasible to employ neural nets to predict individual players' reactions.
|
|
|
Post by ExDeath on Sept 28, 2007 11:31:21 GMT -5
...it Freesoul, you stole my argument.
|
|
Derfel
Ronin Warlock
Did I Do That?
Troublemaker
Posts: 283
|
Post by Derfel on Sept 28, 2007 11:44:51 GMT -5
Man... a Warlocks AI that bluffs? That would TOTALLY rock.
|
|
|
Post by Slartucker on Sept 28, 2007 13:29:08 GMT -5
What we really need is a trash talk AI.
|
|
|
Post by nawglan on Sept 28, 2007 13:36:11 GMT -5
2 Slartucker:
You have something on your shoulder. . . . Hey.. Look over there. *queues up FoD*
|
|
|
Post by Slartucker on Sept 28, 2007 13:44:13 GMT -5
If anyone ever reaches the "Grammatica" title at its new and lofty postcount, I'm going to give them the secondary power of being able to strike down gratuitous abusers of language: those who uses "2" as "to" in a single-thread conversation where even the "to" is redundant.
...which just goes to show, I'd like to see a bot out-trash talk me :-D
|
|