shampoo wrote:
It is a good idea to play against highly skilled computers with uneven conditions due to the technical programming facts that computers will change their styles of play on the fly as the match progresses.
For instance, let’s say there is a program that is around 5 amateur dan level.
If a person with 3k level plays against the 5 amateur dan, the 5 amateur dan will win with the score of +10 or so.
If a person of 1k level plays against the 5 amateur dan, the 5 amateur dan will win with a score of +10 or so.
If a person with 2 amateur dan level plays against the 5 amateur dan, the 5 amateur dan will win with the score of +10 or so.
If a person with 3k level plays against the 5 amateur dan, the 5 amateur dan will win with the score of +10 or so.
If a person with 4 amateur dan level plays against the 5 amateur dan, the 5 amateur dan will win with the score of +10 or so.
Now we can see where this is going.
We can get the gist of this great mechanism of GO computer programming that is analogous to the contempt value and aggressive or defensive modes of chess programming.
For almost all GO playing programs, they are programmed in such a way that the programs will play defensively to maximize the winning chances owing to being defensive is better at winning when it is winning.
That isn't a matter of the programming so much as the underlying algorithm itself acting correctly according to the definition of winning or losing at go. A simple consequence of "margin of victory irrelevant; a win by one point is as good as a win by 20". The exact same situation applies to a "match" sailboat race (a race between two boats, like the finals of the "America's Cup") If you are ahead, the proper strategy is to keep your boat between the opponent and the next mark as this guarantees remaining ahead no matter how the wind might shift. If you are behind, proper strategy is to veer to one side or the other of the direct course. If the boat ahead does not "cover" (what I said before) this gives some chance of pulling the race out if there is a favorable wind shift at the negligible price of losing by a larger margin if the wind doesn't shift or shifts the other way.
You didn't mention the converse behavior. If behind, the MCTS algorithm will play more and more risky moves until they become so hopelessly risky the game is getting worse and worse and eventually the program resigns when calculation gives no chance at all (or below the chance that triggers "resignation"). The same thing tends to happen in the sailboat race. Each time the boat behind veers the boat ahead in "covering" can usually cut the corner somewhat and so be increasing its lead (it doesn't have to veer quite as much as the boat behind did to stay between the boat behind and the next mark).