Check out the Generation 5 website.

P.S. Mods don't like it when you post the same question in different forums. If you post, then find another forum that it better fits into, delete the misplaced post.

[edited by - Vikato on May 29, 2002 11:28:02 PM]

Searching a Chess game state-space is just about within the grasp of a modern PC (in fact, whether you use a basic PC clone or a supercomputer doesn''t really make much difference because we a generally talking about a logarithmic scale).

So typically this is how computer chess algorithms work - just add a good heuristic for ''pruning'' the search.

The branching factor in Chess (basically the number of possible moves given any board position) is about 35. The branching factor for Go is more like 200, meaning searching this game tree is not a viable option, hence a far more interesting problem (in my opinion anyway!).

There is at least a second reason why Go is a more interesting problem than Chess, and that is that the rules for Go are extremely simple compared to Chess and quite often board evaluations can be estimated by a human player based solely on visual cues (as opposed to Chess where complex interactions between pieces need to be resolved in order to evaluate the board).

This is why the best Go computer player can generally play as well as someone just starting to learning to play the game. After a handfull of games though, the human player has usually learned to beat the AI.

quote: Original post by Carrot

There is at least a second reason why Go is a more interesting problem than Chess, and that is that the rules for Go are extremely simple compared to Chess and quite often board evaluations can be estimated by a human player based solely on visual cues (as opposed to Chess where complex interactions between pieces need to be resolved in order to evaluate the board).

Sorry, you are incorrect. In the past I played Go at competition level for a number of years(never got my Dan though ;-() and I can assure you the interactions between stones can become incredibly complex. The wonderful thing about the game is that how one evaluates a board position works on many levels and is almost fractal in nature. This is hard to explain in words. It''s only something you start to realize after you become fairly adept at the game, around 6 kyu''ish (and that takes a fair bit of playing for your average Joe!). Unfortunately a lot of researchers working on the AI for Go do not play this well so they remain unenlightened. I think this is a major factor restricting development of Go AI.

Go, in general, is a very misunderstood game.





Stimulate

I realise that the interactions between the stones can be very complex, but what I meant was that it is much easier for a human player to evaluate a Go board than a chess board given similar levels of competance at both games.

The reason for this is because the interaction between the stones in a game of Go depend on the spatial relationships between the stones (hence the good visual cues). In a game of Chess each piece has its own individual characteristics making visual cues less important (whereas in Go all stones are equal).

It''s much easier to get an overall ''feel'' for the game with a quick glance. In this respect the board evaluation is a little like pattern recognition (which is the point I think you are trying to make). But all this ''feel'' is doing, is directing your attention to several subareas of the board which in turn resolve themselves at an even more detailed level. The final evaluation of the stones can be an extremely complex process. Just as complex as any chess game, if not more so. (but then this whole debate has been done to death by chess and Go players for years) Chess players think chess is more complex. Go players vice-versa. As hardly any good chess players play Go at all well and any great Go players play chess well this debate is likely to run on for centuries.





Stimulate

yeah, thats true, the Go/Chess debate is an old one, but I''m not debating that as such.
I''m comparing the two games within an AI context - Go is much more difficult in this respect.

You seem to thing that I regard Go as a game lacking complexity!
If you look back at any of my posts in this thread I''ve never said anything like that. I said the rules are simpler (they are).

That ''feel'' for the board you talk about really sums up my point.
A computer Go player lacks that intuition you talk about. Now I know you say that it just direct your attention to a general area of the board, but you underestimate that power.

If we could get a computer algorithm could even do this we''d be a much closer to finding a decent AI player! In fact the complex level of individual stone-interactions you talk of is the area where the AI would excell.

From what I've read, most go AI's have attempted to cull out all portions of the board that aren't "relevant" to the game, but this is a tricky thing to do. For instance, a "ladder" move series could continue for ten or so moves, making an empty portion of the board crucial. In order for a computer to recognize this possiblity however, you have to look at the entire board anyways, so the benefit of the culling is made fairly negligable.

Fup, this "feel" you have for the game is courtesy of your large, complex brain, and I challenge you to pick out exactly what it is that you recognize in order to get that feel. Maybe this could be accomplished through a multi-layer neural net, but that raises the whole question of what would constitute the training set. Any thoughts on this?



[edited by - Mordoch Bob on May 31, 2002 7:37:03 PM]