AntiComputerStrategies

   

As of 2012, computers have reached 5d on KGS, 6d on blitz accounts using MCTS and it is an interesting question what their weaknesses are.

Old Discussions

In chess, there used to be well-known strategies for exploiting the weak points of computer players. This seems a bit premature in Go, because computers are so weak: the best anti-computer Go strategy is probably "study until you're stronger than 10 kyu".

Recently I (CharlesSutton) have been playing about 6 kyu over the board, but when I went on KGS, I had trouble with GnuGo, which is ranked about 12k or 13k. This annoyed me, and started me wondering what I should be doing to beat that @#*! program.

Here are my generalizations about computer play (based on little experience, I admit):

  • Computers make strange tenukis, leaving behind groups whose status is unsettled. Often, you can ignore the computer's tenuki and kill the group it left behind. Always responding to the computer's move is a bad idea. (Of course, it is a bad habit to always respond to your opponent's move, no matter who you're playing.)
  • Some programs aren't great at life and death. If you are, then you can probably kill a few of the program's groups. If you're not as good at life and death, you should try anyway, and consider it good practice.

Other suggestions?

nachtrabe: I wrote a bit about this in my blog a few months ago. Here were some additional observations, specific to gnugo:

  • Gnugo is very greedy when it comes to stones it can capture. It will often set up a capture rather than take moves that would net a greater profit elsewhere. Sometimes even ignoring the safety of its own group to go for a capture. You can sometimes even trick it by giving it a capture as a sacrifice and it will spend several moves taking it while you net greater profit around the board.
  • Gnugo likes to play monkey jumps and play hane-on-the-first-line endgame moves very early regardless of their actual value. If you can ignore these and they arenīt really sente, do so.
  • Gnugo almost never seems to play strategic forcing moves, such as peeps, or strategic sacrificial moves, such as probes.
  • It has trouble handling any situation where the opponent feints or otherwise plays in a "two-handed" manner.
  • Sometimes a player can get a feel for the areas that the computer is looking at. Certain invasions it just never seems to consider.
  • The computer's knowledge of joseki is very brittle: it often plays the wrong joseki and if a move isn't in its library chances are good it won't be able to come up with the right continuation. This can give an edge in corner sequences.

Andreas Teckentrup

  • If You play a program often enough, You can find sequences that the program cant handle. See GnugoNosekis

Alex: The most general rule for playing computers is that they are reasonably strong tactically and very weak strategically. A good general principle, then, is to stick to what you know and avoid complicated fights with them. If you settle positions quickly, you bring the emphasis back to deciding on a whole board strategy, which the computer is weakest at.

Incarlight: One thing I've noticed about computer opponents is that a 3-3 invasion always is a good sign. If the computer invades at the 3-3 point it seems to usually be because it knows it's many points behind. But even so, the computer sometimes manages to mess up completely, leaving a dead group in the corner. Or as I've seen once, abandon the invasion in order to connect one of the invading stones to a live group, thereby abandoning the corner completely. (In the end I think it failed both to connect and to create life, if I remember correctly.)

And some computer opponents have a stange idea about life & death. Sometimes it will spend three or four moves capturing a dead group while you take points elsewhere.

hk: GnugoNosekis was a really interesting page, but GnuGo is a traditional program relying a lot on hard-coded knowledge - at least a lot compared to the modern UCT programs. I understand UCT programs struggle with the situations called Semeai, and in general with situations where a lot of "apparently bad" moves end up in a good one. For those of you who play it, do you know many MogoNosekis, or are these programs more resistant to anti-computer strategies?

I don't count playing the windows version of Mogo (30% slower, no multicore, crashes) or playing with very low time limits/poor processors as fair anti computer strategies, although I understand they work!


This is a copy of the living page "AntiComputerStrategies" at Sensei's Library.
(OC) 2014 the Authors, published under the OpenContent License V1.0.
[Welcome to Sensei's Library!]
StartingPoints
ReferenceSection
About