Life In 19x19

Krama,

Bonobo wrote:Krama,

Krama wrote:

Javaness2 wrote:How many times do you intend to ask that?

I guess until they get an answer.

I guess you haven’t witnessed how many times “they” have asked the same question, in different places, AND was given hints about where to begin

hehe

tydsh added a stronger DCNN only version, darkfores2. It's currently 3d(!). While not as high as the strongest bots, I'm very impressed that a neural net with no search whatsoever can get rated so high.

yoyoma wrote:tydsh added a stronger DCNN only version, darkfores2. It's currently 3d(!). While not as high as the strongest bots, I'm very impressed that a neural net with no search whatsoever can get rated so high.

It's kind of a digital reincarnation of TMark Hall! As I understand it, darkforest2 has been fed the GoGoD games database, which TMark played through and recorded, and - like him - plays very quickly, based solely on shape. And plays at a similar grade to his long-standing 3 dan.

Freewheelin' wrote:

yoyoma wrote:tydsh added a stronger DCNN only version, darkfores2. It's currently 3d(!). While not as high as the strongest bots, I'm very impressed that a neural net with no search whatsoever can get rated so high.

It's kind of a digital reincarnation of TMark Hall! As I understand it, darkforest2 has been fed the GoGoD games database, which TMark played through and recorded, and - like him - plays very quickly, based solely on shape. And plays at a similar grade to his long-standing 3 dan.

That's a nice thought, but T Mark was 4 dan (though his rating did drop a bit in the last few years).

I think winning against this DCNN is quite easy. Play unusual shapes and start fights. Since it can't calculate it will probably fail.

Krama wrote:I think winning against this DCNN is quite easy. Play unusual shapes and start fights. Since it can't calculate it will probably fail.

Looks like you should just go play it then. Let me know when you do - I'd love to watch!

I just noticed in one game that the network can't even read ladders. So an easy solution would be to start a ladder that works for you.

You are making some assumptions (mistake) about how a neural net, or a MCTS program for that matter, works.

MCTS -- in its purest form, very simple but too slow. And THAT is the "ladder problem", not that it doesn't read ladders, because the pure form isn't reading ANYTHING. It is an "evaluator", is move "a" better than move "b" based on the concept that if ENOUGH random games are played form that point on by two EQUAL players (equally good or equally bad) and a higher precentage of games is won following move "a" than following move "b" then "a" is better than "b". The "ladder problem" (and similar things)with this simple form is that "enough would be enormous. the "equally bad" players have rules to guide them, not random play << eg: recognize an "ladder" and get the laddering moves into the tree >>

So how MCTS performs is a tradeoff on these special rules controlling how moves get added to the tree to speed things up while hopefully not pruning away the right best move.

Neural nets are TRAINED. What they are is something that computes a function. Given an input, produce an output (position => next move or perhaps small set of moves to be fed to a MCTS evaluator). The large database of games is the TRAINING DATA. Think of this as a multi-dimentional array of cell values with connections between some cells. During training, cell values are altered and how values transfer through connection according to whether it gave a right answer (reward) or a wrong one (give those values and connection coefficients small random changes until helps with the answer for that position without hurting the results for the other positions; then on to the next training position. There's more to this. Ever so often, a random small disturbance to all values and connections, and retrain (that's called "annealing", necessary to get off false summits).

The point I am making is that the neural net doesn't contain that database of game records in their original form. What has been learned has been internalized in the final version of cell values and connection coefficients << note: training the neural net is time consuming (computationally intensive) but once you have a trained neural net you can "clone it" instead of having to train a copy (simply initialize to all those cell values and connection coefficients).

What is interesting about neural nets is the trainer doesn't have to know how to it (evaluate the function) but just be able to tell right answer from wrong (the assumption here is that the record of a game between high level experts represent almost all correct moves).

Understand? Sticking to ladders for example, doesn't have to "know" about ladders. In the position where a ladder COULD have been a factor, the expert didn't play in a certain way or perhaps played a move that was a ladder breaker preventing the ladder from materializing. The neural net doesn't "know" anything. Look, at THIS level (the level of the internals in your brain) you don't know anything either. Just signals going between a net of neurons.

BEFORE HAVING BEEN TRAINED (to some specific task) there is no difference between neural nets of the same complexity. In other words, the neural net program is "general purpose". Only the DATA stored during the training process would distinguish between one to play go, to play chess, or drive a car (I know Carnegie Mellon had a neural net doing that quite a ways back)

I looked at its games, and while there were a few fighting/reading sequences that seemed dodgy to me, there were not many. I'm at least 4 stones weaker, so I'm sure a dan player would see more holes, but I don't know if the bot has any obvious weaknesses a kyu player can exploit.

yoyoma wrote:

DCNN = Deep Convolutional Neural Network -- most commonly used in image recognition. So it is good at recognizing shapes/patterns on the go board.

Misunderstanding the generality of a DCNN. They are evaluating a function and the function could be anything. Useful for when those training them DON'T know how to do it, just can tell right result from wrong.

Image recognition a good example. Say it's face recognition. The trainers know for each face what the correct name is and in this case they can themselves learn to perform this task (assigning name to face) BUT they don't really know how they do that. Nor is it necessary that the be able to do it. All they need is a good supply of images for each face to be recognized, each correctly labeled. A DCNN can then be trained to recognize the faces that but NOT because "good at recognizing shapes/patterns". You have just read in a "how accomplished" but that doesn't make sense when a neural net is involved. They "just work" (once the cell values and inter-cell connection coefficients have been properly adjusted, and that adjustment is by random trial and error. Under training the originally blank neural net evolves to one that correctly evaluates the desired function.)

Think of biological evolution. No PLAN involved << sorry if I just offended "intelligent design" folks; but the real thing is more elegant and robust than "by design" >>

Off topic:

Interesting consequences of having an AI stronger that humans : intrinsic time-independant ratings.

http://www.cse.buffalo.edu/~regan/paper ... eHa11c.pdf

I will now retract my claim. The DCNN is quite strong (I still didn't play it)

I was able to win against 5/6 dans on KGS and plays on a stable 3d - 4d

However there is another theory that I have, the bot actually plays bad because it plays amateurs and not pros. Remember that it was trained on pro games.

It would be interesting to see it play against a pro but without pro knowing it is a bot (thus preventing the pro from playing overplays thinking the computer will fail to win the fight)

The NN plays quite good in the opening but it is the middle game that the most mistakes are made.

I have watched it play almost pro like only to blunder groups for no reason.

Mark Zuckerberg (Facebook boss) just made a post about their Go Ai with a link to that paper, I wonder if that'll get more people interested in and playing Go?