Re: Computer version of Risk
- Posted by Anders Eurenius <c96aes at OXE.CS.UMU.SE> Jul 04, 1997
- 717 views
FIRST OF ALL: I feel pretty stupid arguing with a pro, but since it's turning out to be really interesting... > It kept track of every move of every piece then analyzed the board after > each move. It would ask each piece what it now could do or not, then Ok. I think the comment is a bit odd... "Ask each piece"? Is this any different from listing of all possible moves? Is the "valuation" for these= moves "deep" or shallow (no tree)? > prioritize each possibility then advise the player of his options now. Sift though the tree for highlights. Ok. > The player could view every move for every piece and it would tell the > advantages and disadvantages of every move. Tactically, how did the Whoa. Hold it. Now we've got a hardcoded tactics expert scoping out the tr= ee. Is it a new-and-improved valuation function? It observes regions, position= s, complementing/canceling "threat bubbles", trades, and maybe some other contructs too? Oh, yeah, It's real adaptive too? It sees all kinds kinky s= tuff? (Am I looking in the right direction?) > players situation change as a result of that move? Same evaluation for resulting board. Ok. > You miss the beauty of chess here. It's not winning the game, it's HOW (How?! Wow, I'll have to meditate on that one. I can relate to the beauty-= thing, but I can't connect it to chess...) > you do it. It ceases to be AI if it's just going for points. There are Yes. It will fail a Turing-test. (Unlike IBMs code/machine) > brilliant strategies that take 20 moves to accomplish, those 20 moves ma= y > not be the "optimal" point wise, but they win the game faster than the > point wise moves. The freaky part about max-min, is that it *will* choose the niftier strate= gy, *if it sees it*, that is, searches deep enough. Is your emphasis on playin= g like the humans (suckers!) or giving a hard fight? I don't a lot about che= ss, but I figure there's got to be about as many brilliant strategies as there= are players right? (In addition to which, they too, count a mind-warping numbe= r of moves ahead...) How do you make the computer use one of them? Some kind of= hardcoding? When do you abandon a master-plan like that? As for playing like a human, well, I have a real problem with believing th= at any chess-player is really human... ? > I'm not talking about a random-factor that every once in awhile does > something different for the hell of it! I'm talking about a deliberate, > calculated action that is in the character's game plan that may or not b= e > in his best interest. We all try something stupid when we play, > sometimes it pays off bigtime, sometimes we look really stupid. Hmm, yeah. I wrote the original post because I felt that what you were describing (in terms of algorithm) was total free-hand spaghetti. It's oka= y to make spaghetti, but you should know what you're doing, otherwise it will b= e *just* that, and nothing more... My point is to first write a shake-and-ba= ke by-the-book algorithm, and try to get it bent later... (If at all) > Pretty much. Teach the computer player how to play, then what you do wh= en > you play and tell it how you choose your move based on whatever info is > required, then let it play. As you play against it, you adjust the "how > you choose" part until it plays better than you, then get other people t= o > play against it and get feedback. By maximizing score, I mean the kind awarded by the AI-valuation, not the = game rules, and they may be very different... (So our views on the resulting mo= ves aren't all that different. It's your attitude on the code I really have a = beef with...) If the computer beats me, and gives me some attitude doing it, it= 's Ok. But giving me attitude and getting stepped on is *really* *losing*. > > Example to prove what? > > (All I see is the results of a too simplistic view on AI) > > Turn and shoot is the "optimal" way to win, but not the best way. > Your method is go for the points, you kill by pointing at the target and > shooting, not by flying into formation in your enemy's sights for awhile= , > taking hits then sacrificing 1/3 of your attack force to give the others= a > possible better shot at the target. You're assuming that the AI is /very/ short sighted! This is not neccesari= ly so. max-min will happily sacrifice anything, if it sees the gain. (this is= the hard part though) > > Like this isn't voodoo...! "Learn". Yeah, sure boss, we'll put that in= before > > lunch... > > It ain't voodoo. It's keeping track of what you are doing, what you do (If you've know how it works, it is never voodoo, that's, like, the definition...) > often, and choosing another route based on that info. It's the same > stuff. If I play 10 games against you and in every game you go for afri= ca > with reckless abandon and leave other things helpless, I will try and > capitalize on that weakness in your strategy. It's just tables. It's > just telling the computer player something, and what to do with that > something he knows. Yeah. And what do you watch for? How subtile does it get? Going for africa= every time is a bit big... ? > If I know that as soon as one ship breaks formation you follow it most o= f > the time I will look to see if you do it this time and have the other > ships do something different in response. In practice it's more like on= e > ship tells the others that it's being followed and the others try and > sneak up from behind. If I know a particular strategy isn't effective > against you, I'll try something different. Go to plan B. And keep a hardwired set of plans/strategies. Doesn't sound very bright to= me... If it is really intelligent, it surprises it's creator. And that wou= ld IMHO definately requires some voodoo... > > The max-min would prolly keep its playing style, but controlling depth= and > > random-factor it would probably still be tricky to beat... > > I'm sure it would be, but is it "smart?" Would it be like playing again= st > a real person? Do we want to play against a human we *know* we can beat? Seriously, I hav= e a hard time accepting human traits as "smart". The sharper the players, the = less quirks, preconceptions, superstitions, habits, whatever... I would think t= hat at a lower difficulty, the AI should have more "personality", and less hig= her up... > > (Oh yeah: And be careful with equal signs around here...) > why? Some mailer or some server somwhere is puking out "=3D3D"... Pretty #&$=A4= % random! So equal signs get filtered out by eyeballs when reading... > Michael Packard Anders ------------------------------------------------------------------- Anders Eurenius <c96aes at cs.umu.se> ICQ UIN:1453793 Computer Science/Engineering student at the university of Umeaa -------------------------------------------------------------------