I am not sure how they work now but the older chess programs were not AI in the sense of the modern AI. Those chess programs could look at trees of possible move and countermove, intelligently pruned to limit the quantity, and then evaluate end states X number of moves down that tree, assign those end states values based on what really s…
I am not sure how they work now but the older chess programs were not AI in the sense of the modern AI. Those chess programs could look at trees of possible move and countermove, intelligently pruned to limit the quantity, and then evaluate end states X number of moves down that tree, assign those end states values based on what really smart chess players gave it as a template and then choose moves.
You simply cannot do that for Go. The board is 19x19. Each move only takes off one space and, if there are stones captured, may add more spaces that can be moved to. You cannot brute force look at trees. Each move branches off into too many possibilities and you need a average of the aggregate impact of each of those trees. The calculations involved are too many. There are something like 10^700 possible Go games for every atom in the universe.
The value of an individual move is different for early, mid and late game and the strategy requires a global awareness of how a move will impact adjacent areas that are developing differently. As far as I understand it, the Go AI were trained by showing them games and having them play games. There was no way to tell them what is good playing or strategy. They learned it by figuring out how to win. That is a true AI and on a completely different level from the older chess programs.
Very interesting, thanks. So chess is much more a closed system? - I do know that when Big Blue famously beat Kasparov (I believe), the IBM engineers did not (of course) intervene during the matches but they did tweak and optimize their software after every match - it really was one against Big Blue+4 or 5 humans, i.e. computer-augmented chess. Your point is that procedure is not even remotely possible in the GO case - humans have nothing to add and could not remotely conceive how to even attempt to do so.
OK smarty pants - but we humans are un-matched in merging our vehicles into a stream of fast traveling traffic on highways; the vehicle auto pilots (I have heard; well, 2,3 years back) are still pretty hopeless at that task.
I am not sure how they work now but the older chess programs were not AI in the sense of the modern AI. Those chess programs could look at trees of possible move and countermove, intelligently pruned to limit the quantity, and then evaluate end states X number of moves down that tree, assign those end states values based on what really smart chess players gave it as a template and then choose moves.
You simply cannot do that for Go. The board is 19x19. Each move only takes off one space and, if there are stones captured, may add more spaces that can be moved to. You cannot brute force look at trees. Each move branches off into too many possibilities and you need a average of the aggregate impact of each of those trees. The calculations involved are too many. There are something like 10^700 possible Go games for every atom in the universe.
The value of an individual move is different for early, mid and late game and the strategy requires a global awareness of how a move will impact adjacent areas that are developing differently. As far as I understand it, the Go AI were trained by showing them games and having them play games. There was no way to tell them what is good playing or strategy. They learned it by figuring out how to win. That is a true AI and on a completely different level from the older chess programs.
Very interesting, thanks. So chess is much more a closed system? - I do know that when Big Blue famously beat Kasparov (I believe), the IBM engineers did not (of course) intervene during the matches but they did tweak and optimize their software after every match - it really was one against Big Blue+4 or 5 humans, i.e. computer-augmented chess. Your point is that procedure is not even remotely possible in the GO case - humans have nothing to add and could not remotely conceive how to even attempt to do so.
OK smarty pants - but we humans are un-matched in merging our vehicles into a stream of fast traveling traffic on highways; the vehicle auto pilots (I have heard; well, 2,3 years back) are still pretty hopeless at that task.