Not really surprising as there are far fewer things to think about in the second half of the game.

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

There would be a truth in there, but maybe only a partial one. Pro time-management, if graphed (does anyone do this?), and smoothed out, would probably be consistent with the thought.

But, but, ... come the endgame and definite one-point errors can be glaring, to a pro. And I don't suppose one gets away with saying they also actually lie on the surface.

Urrgh. One thing that a strong AI might cast a light on is "the moment when the board really can be treated as a sum of sub-boards", coupled only weakly by ko threats. I'm not quite sure how. There is the traditional oyose, which I suppose starts when both players can be sure that a big group isn't going to die. Then there is the "phase transition" to sums of sub-boards.

Maybe AlphaGo has an edge when those two points in the game are widely separated? Generalising from one instance, or fewer, which we all enjoy, that might be the lesson from Game 1.

I was not thinking at all about the pro when I made my comment. I was thinking of the computer. The more moves left before the game ends the much faster the potential game tree grows. I know that AlphaGo is using nets and not tree searches but it still seems that there will be far more opportunities for the computer to play better during the latter part of the game.

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

I think they could release the final policy and valuation neural nets without giving away general intellectual property. I think the business value here is in the process, not the product.

On the other hand, their Nature paper actually provides an extensive description of how they produced this AI including details on the parameter values that they used. Maybe someone with more interest in the Go community could make a version that would be useful for learning!

I see some significant potential for manipulation here. If AlphaGo is not made public afterwards (by playing other pro's too or in whatever way), I could easily see Google to bribe Sedol into not giving his very best, so that Google gets all the publicity. That is always a danger when a lot of fame/money is involved and we have a one-to-one-fight (ask boxing fans). Just as a reminder.

I think this discussion is handicapped by too many of being confused about what a neural net IS as compared to other sorts of programs.

Step back for a moment. Think of a human brain. Think of a newborn child. The child's brain starts out not knowing any human language. But in its first year of life it will be LEARNING one. The brain of the child didn't start out different before it has learned say French vs having learned Mandarin. It was the changes during the learning process that made the difference.

Well the software neural net is like that. In the beginning a program which simulates a lot of nodes and defines how these nodes communicate to their neighbors, and for each of those communication paths a value that modifies (amplifies or attenuates) the signal sent along that path. The ONLY thing that changes during training is those values. In the beginning, the neural net can do nothing. After having been trained, it can evaluate some function (in this case, input "a board position"; output "a move").

The SAME neural net could instead have been taught to play chess or do anything else. The neural net program itself is knowledge neutral. It's the training process that makes the difference. But of course, having trained a neural net (accumulated all those values) could easily "clone" the knowledge into an identical (but empty of knowledge) neural net. Just copy those values. But note, VERY IMPORTANT, that only would work for an IDENTICAL neural net (not a smaller version.

So no, CAN'T do a "dumbed down version" for small machines. This is a very big neural net, lot of nodes, lots of "signals" going between nodes, takes a very powerful machine to get all that done within the allowed real time.

What the "dumbed down" question is really asking is could there be built a smaller neural net that could also be taught to play go and how strong would that be. Open question. What we have here is only showing us the upper bound.

A couple of other points. A certain amount of randomization is involved in the training process. So if (instead of copying in the values) we trained that second neural net using the same training data we would end up with a (trained) neural net able to play at the same strength BUT the set of values would in general be different!

In terms of "intellectual property" law, this is probably new territory. There is no particular value in the implementation details of the neural net per se. The real costs (and resulting value) come from the training process.

Interesting. So how much hard drive space would such a big neural net take? I had hoped it would be like any other program.

The hard drive space to store the code of the program when it is not running is not important. Even if the code is big, and storing all the various parameters from within the network (which get tuned by the training) is big, hard disks are dirt cheap. The expensive factor is the memory and processing power required to run the program. A program whose source/compiled/whatever code is just 1 KB can contain an instruction saying "allocate 100 TB of memory".

A DeepMind programmer created a chess engine based on deep learning a few months ago. It runs smoothly on a cheap laptop. I would expect the same to be true of AlphaGo, even if its performance on such a machine will clearly not be pro level. At the very least, the version without lookahead shouldn't require many resources to run (and not much time, either), and it's allegedly as good as the previous state of the art.

Are you absolutely certain about that? Since here somebody is arguing that some ANNs can be compressed to save space without sacrificing accuracy. See for example http://arxiv.org/abs/1510.00149.

AlphaGo was playing a teaching game against the human, thus it made the obvious pro mistakes in order to see if the human is strong enough.

I don't expect it to be able to do a teaching game with Lee Sedol but winning should be easy for the AI overlord.

Hail our new robotic overlords!

We're getting nowhere.

This NOT a data problem. This is an intensive computation problem. "All those parameters" wouldn't take a great deal of space.

Look, someday, if and when neural nets become greatly used, we will have HARDWARE to implement the cells, and so neural nets where the nodes are all doing their thing in parallel (instead of the process being simulated by a linear process). Could be orders of magnitude faster.

Your brain cells in the neural net that is "you" are receiving and sending (to neighbors) in times that might be in thousandths of a second (not billionths as in our modern electronic devices). But there are lots of them doing that simultaneously. So in half a second real time, analyzes the visual image determining that is a ball coming your way, and you begin to move to catch it.

@mikenovack: can u recommend some book or link about neural networks and how they work in the brain resp. computer?

This is incorrect. The topology and other characteristics of the network is chosen for a particular problem. The dimensions of these networks were built specifically for the 19x19 dimensions of a go board and for the tri-state nature of the points. Furthermore, there must be *some* explicit knowledge built into the system as a whole - it is expected and OK that the network will suggest sub-optimal moves, but it cannot suggest illegal moves (e.g. it must know the ko rule).

In theory, you could wire in chess games in a really hacky way, but the network probably would work very poorly or not at all. It would be the electronic equivalent of traumatic brain injury.


Such highly-parallel hardware exists and is already in most PCs - GPUs. And neural networks are already in broad use. They're the reason that speech recognition on your phone works quite well these days but was quite terrible 5 years ago.

_________________
KGS 4 kyu - Game Archive - Keyboard Otaku

Free online textbook:
http://neuralnetworksanddeeplearning.com/

They'll be providing some sponsorship for the London Open 2016.

Well, the AlphaGo team did actually train a "compressed" neural net for the move prediction policy engine. It sacrificed a significant amount of accuracy (57% -> 24%), but sped the computation up by over 1000 times (3ms -> 2µs). Enough to be worth it for evaluating the Monte Carlo rollout.

That's not a compressed neural network but just a frequency-based specialized pattern classifier.

_________________
Go programmer and researcher: http://pasky.or.cz/~pasky/go/
EGF 1921, KGS ~1d and getting weaker

Ah, my mistake. I see the paper you referenced is specifically referring to compressing an existing neural network rather than training a smaller network.

incredible, i didn't think it would happen for another 5 years. an 8-9 Dan computer go program.