This is what I was also thinking about.

Since we can say the network can guess the correct move (one played by pros) 55% of the time. Then why not order the network to give us let's say 5 moves. If the correct move is found 55% of the time in 5 moves there is a 95% chance that one of those moves would be played by a pro. With 10 moves selected it goes into 99.75% but it would probably be an overkill and slow the whole process down.

So asking a network to give us top 5 moves, we can be pretty certain that we get a correct move. However if we repeat this process with let's say 125 turns (if we take that an average game lasts 250 moves) then something bad happens. In 6 turns out of those 125 you get all wrong moves. However if you increase the number of moves you get from the network to 10 then you can pretty much be certain that you will get at least one of those 10 correct through the whole game. However I still have so much things on my mind that can prove this method not so good but I just can't bother to write another walls of text. Maybe in the next couple of days I find some time to write some other ideas on how to improve go AI.

They deliver: viewtopic.php?f=9&p=198123#p198123

Time to pay up palapiku?

You are missing something (misunderstanding "right move" and "wrong move")

There (usually) isn't ONE move made by pros (for most of the moves in a game). A neural net is trained on the records of actual games played by pros. When you say 55% that's the move made in that situation by that pro. But when it returns some different move, that's the wrong prediction (of what the pro did in that game) but NOT necessarily the wrong move. Not necessarily a BAD move. A little thought and you would realize that on at least move in each game between top pros a mistake was made (somebody lost, didn't they). So even when the neural net returned the "right move" (the move the pro made) that might actually be a "wrong move" (in terms of the game).

So a modification that returned five moves (that set of five moves) might have a high probability for the set to contain the best move. MCTS could be applied to the five.

No need to be condescending.

_________________
Occupy Babel!

I feel heartwarming, because you know what, Demis Hassabis, your go disciple, is honest about his word, he deliver it at the beginning of 2016, and I hope he'll close the case (by beating the current champ, not Lee Sedol) in 2016 as well.

Thanks for the kind words.

If AlphaGo can establish some sort of claim to be in the 4p to 7p bracket, then the whole business of computer go is clearly changed. One victory against Lee Sedol that wasn't decided by a blunder might do that.

So far I think we know the program has "talent", and is worth watching from a spectator's point of view.

Actually, I'm glad that I didn't read this topic back in 2015, cause I might reply with something like 'I don't believe you'

Some food for thought, Google also buy another frontier company in making robot with legs, Boston Dynamic
and what if you combine the best software from Deepmind and the best hardware from Boston Dynamic and let it challenge human in the game of humanity... yes, I mean football/soccer! The idea and condition are very rough right now but you know it's gonna happen at some point, hope it's in my lifetime though.

I guess "robot with wheels or wings" is more à propos. DeepMind will turn to 3D games, according to Demis Hassabis in a video interview. So, cars that drive themselves, drones ... all that mundane stuff.

By the way, do you know how good Hassabis is in go? He states that he's not a strong go player, but a very strong chess player though. I'm guessing he's in single digit kyu?

Based on what I know about AI neural networks I doubt that this program can explain in conceptual terms why the moves it chooses are good.

When we played a series of handicap games, he got up to about 4k rather quickly. I coached him through a couple of online games. After that, I don't think he played much at all, seriously. So assume he's a mid-range club player - safest.

But he has an eye for games, and has been involved with go recently, so he might well have improved somewhat.

This program, no. But recurrent and convolutional neural networks are also used for natural language processing so...

_________________
KGS 4 kyu - Game Archive - Keyboard Otaku

Whilst being able to explain things conceptually makes a great teacher, it's not required for a bot to give useful teaching input. For example I gleaned some useful things from CrazyStone's analysis of my game here, notably some of the fighting sequences and trades I had not considered around the lower right corner from move 66. Whether I can incorporate those specific sequences and ideas into my general playing strength is less clear, but I don't think that's because the CrazyStone was unable to say "Broaden your mind and look at the bigger picture and try to consider more possibilities for trades and exchanges; is that move you think is sente really sente or can you make some exchanges first?" as I was able to create that conceptual lesson in natural language using my own intelligence applied to its analysis.

I think the question "can the bot explain" (doubt that this is possible) is partly a misunderstanding about different parts of a program.

"Can" is not the same as "does". And "provided by the same algorithm" not the same as "provided by ANY possible function".

In other words, a particular bot (currently) might simply return the next move. It DOESN'T offer a function "and explain the purpose behind that move". But this does not mean the program could not have been provided with a facility "explain the go reasons behind that move". Understand? The program might not have used those go reasons, comparing with the go reasons behind alternative moves and choosing which was the better move based upon most compelling go reasons. But that does no mean that GIVEN a move an AI could not be devised that explained "what does this move do".

To give a concrete example, take MFOG12. Its higher levels use MCTS to select the move to be made, but you can turn on "why" and using the AI which plays the lower levels, it can list the "go things" that move accomplishes << even though that AI can't evaluate which set of "go things" better than those associated with an alternative move at much better than 5-6 kyu level >> Being able to assemble a set of "go things" for each potential move is NOT the same problem as then being able to judge which set best (and so which of those alternative moves best).

So ........ to answer the initial question. A bot using a neuural net to PLAY go could be provided with a function that explained "what is that move good for" using any of a variety of techniques, and as the previous person pointed out, might be another neural net trained to THAT task.

http://www.nature.com/nature/journal/v5 ... 961-f5.jpg (the picture was too big to attach to the post)

This is from the paper. The network doesn't "explain" it's reasoning in natural language, but you can see how it valued different options. There's some interesting stuff there, I especially like the "direction of play" at the top: White split two 4-4 corners: which side should black approach from? In section F you can also see the next 26 moves AlphaGo sees as most likely.

If you had access to the engine, you could force it to play a move you considered, and see what it viewed as the most plausible continuation from there.

In some ways, I think this is actually better pedagogically. You have to tinker and analyze with the results to pull out lessons, but hard won lessons seem to stick better anyways.

I just watch he give an interview with korean press and he said that he himself is 1 dan!

Fair enough: does destroy my story that out of my five "serious" pupils he was the only kyu player.

This is a few offtopic:

What AI is for sale to play against it? I think that we can play against it to train. I can win GNU 3.8

SL has a list of go-playing software. It might be a bit outdated at places, but it's still a good starting point. You can find some discussion of specific products here in the Computer Go subforum.