xed_over wrote:aw..phooy...
some of you guys put too much emphasis on a bot's weaknesses, the hardware its running on, or that it must be so far above you to be of any use.
AlphaGo is not that much more above pros, yet they are all eager to see what they can learn from it to improve their own game already.
to the OP... find a bot... have fun. Learn what you can from whomever you can.
The result of over generalizing? Not looking carefully at what is being said?
Look, it depends VERY MUCH on the algorithms the program is using and that is closely related to strength levels and strength levels are in fact what strength levels can be achieved related to hardware power.
a) For strength levels up to about 6 kyu the programs almost certainly "go knowledge" based AI's. These don't need a great deal of computer power to be able to play with reasonable time controls. But they tend to be systematic and not aggressive. Many of them will make the same move in the same situation all the time.
It is THESE you can learn bad habits from unless set at a level well above your own.
b) For strength levels above about 2-3 kyu up to perhaps 3 dan (with a strong home computer) and up perhaps 6-7 dan (with a machine well above that in power) the MCTS based programs can be used. But they cannot be easily* weakened below that and this is more the underlying algorithm than the implementation (it would get erratic). But these do not have SYSTEMATIC weakness, you won't learn bad habits, won't ALWAYS play the same, so you can use them set to your own level.
When people say "hardware matters" it is THESE that they are talking about. Needs to have enough crunch to do enough playouts within the time control.
c) AlphaGo --- here we are talking about an approach and a program that requires a machine at least an order of magnitude more powerful yet, say two orders of magnitude more than a powerful home machine. As an neural net trained on moves made by high ranking professionals, of course its analysis would be respected by high ranking professionals.
There is no point in discussing this one as a bot you could use.
* In theory could have a MCTS program that could play weaker by combining "a" with "b". Picture a sort of tag team play, randomly selecting whether the "a" player or the "b" player made the next move, strength level set by adjusting that probability. Sort of like a weaker player who sometimes makes brilliant moves. AFAIK, nobody has implemented something like that.