AlphaGo Teach – Find new and creative ways of playing Go | Hacker News Reader

AlphaGo Teach – Find new and creative ways of playing Go

alphagoteach.deepmind.com

99 pointsJach23 comments

Many comments seem to miss the real significance of this.

This is not a joseki dictionary, and it doesn't play 50+ moves for you. This is AlphaGo Master's explanation of how it sees the opening. Humans have a very good understanding of endgames, and we might understand opening in a single corner, but when it comes to whole board knowledge, we are way worse than AlphaGo is.

As it happens, AlphaGo disagrees with human play a whole lot, just in this subset. There's well known fuseki strategies played by professionals that AlphaGo thinks are downright terrible, and it shows lines of how it would play against them.

Sure, it'd be nice to have even more moves and alternatives, but this is really nice, as it helps us in the parts of the game where we are the weakest.

Compare this to what Alpha Zero is teaching us about chess: It has interesting opinions regarding a few openings (Goodbye French Defense) but ultimately it's saying that our opening knowledge is very sound, but that it will not just crush humans, but traditional chess programs in the middle game.

When an AI is demonstrably so far beyond human ability that to even comprehend it requires the AI to help explain itself, you start to take Singularity futurists a bit more seriously.

This doesn't make sense. Computers do stuff that people don't "comprehend" all the time. In fact, in the past, one of the fundamental appeals and drives behind AI was to make systems that could explain themselves, not requiring complex analysis and experiments to understand what's going on. (And vice vera, systems that could understand humans without programming.)

> make systems that could explain themselves

Yes, but currently the systems we are building have no sense of will or agency.

If we succeed in building systems that do have more complex objectives, then how can we be sure that the human and machine's objectives are aligned? How can you be sure the system is explaining itself to you in an honest way?

Note that there were two different goals in building computers and software: 1) building things that have full control over and 2) building things that are more and more clever and useful.

In the limit, those two goals can come into sharp conflict.

Is there a way to do anything other than just run through this single opening set of moves? Not a lot of exploration here, or ability to 'find new and creative ways of playing go' Or am I missing something?

There are plenty of branches, but it's true there is no way to add your own for AlphaGo's evaluation. The main function of the tool is to see AlphaGo's appraisal of common opening lines from professional games, plus some of AlphaGo's own suggestions.

click on the circles instead of the forward button

Correct, there is only one branch

There are lots of branches. If you don't understand go strategy then you're not going to get much out of this

I've played thousands of games of go. I understand go strategy. I don't understand the interface.

I got super excited for a second...I thought this was Google releasing a working "virtual teacher" where it teaches different strategies of Go and allows you to play against it with different level of difficulty.

I think it is comparable in function to Kogo's Joseki Dictionary, curated by Alexander Dinerchtein (3p). e.g: http://eidogo.com/

Just that it covers the entire fuseki rather than just a corner.

As you get deeper into a branch it doesn't provide many variations, nor it explains why a variation is better than another. Joseki dictionaries sometimes will tell you one move is a mistake and show you how to punish it.

I don't play Go at all, but I would love to learn from a self-adjusting AlphaGo tuned to win (say) %50 of the time. I feel like a whole new generation of players could be born from a "from scratch" approach.

I missed the part where it said the win percentages are relative to black winning (as opposed the the currently playing side), and thought alphago was saying the best response to black opening with both 4-4 and 4-3 was tengen...

Oh this is cool, is it basically giving you a view of what alphago thinks the win % is for a given board state while you're playing? With undo that's awesome.

what you're talking about would basically mean a full free online version of AlphaGo, with what would appear to be instant next move computation.

Saddly, it's not.

no, it's a very small precomputed tree of moves. you can only click on an increasingly small subset of tiles after each move

Damn :( Thanks for clarifying though

It would be sweet to play a whole game against AlphaGo at some point. Is anyone aware of when we'll be able to do that?

There's a project called Leela Zero on Guthub which appears to be Alpha Zero without it's neural network weights. I doubt it'll get the proper weights any time soon though.

Surprising that it thinks the winrate for black is so low. I wonder what it would think of a komi of 5.5 instead of 6.5?

Is it just me, or does the board go asymmetric way too fast?

This may be a UI issue...

Clicking on the circles seems to do nothing in Google Chrome.

Reply on news.ycombinator.com