Facebook's Open-Source Go Bot Can Now Beat Professional Players (techcrunch.com) 44
Google's DeepMind isn't the only team working to defeat professional Go players with artificial intelligence. At Facebook's F8 developer conference today, the company announced a Go bot of its own that has now achieved professional status after winning all 14 games it played against a group of top 30 human Go players. TechCrunch reports: "We salute our friends at DeepMind for doing awesome work," Facebook CTO Mike Schroepfer said in today's keynote. "But we wondered: Are there some unanswered questions? What else can you apply these tools to." As Facebook notes in a blog post today, the DeepMind model itself also remains under wraps. In contrast, Facebook has open-sourced its bot. "To make this work both reproducible and available to AI researchers around the world, we created an open source Go bot, called ELF OpenGo, that performs well enough to answer some of the key questions unanswered by AlphaGo," the team writes today. Facebook's AI Research group is also developing a StarCraft bot that it too plans to open source.
I see a lot of C++ code in the repo! (Score:1)
I see a lot of C++ code in their repo. I think it's pretty telling that even in 2018 it turns out that C++ is still the best language for writing new code. This just goes to show how badly modern languages like Haskell, Rust and Nim have failed to surpass C++.
Re: (Score:2, Funny)
That's because C++ viciously exploits its privilege to oppress those other languages, which all prioritize social justice and welfare over petty issues like maintainability, performance and portability.
Re: I see a lot of C++ code in the repo! (Score:2)
Re: (Score:2)
No. It goes to show how it's still the best for 'some' projects.
For other tasks, C# or Java or something else would be equally fast but many times easier to develop, easier to test and easier to maintain. It all depends on the context of the project. there is no single 'teh best' (typo intentional) programming language.
#DeleteFacebook (Score:1)
So can LeelaZero (Score:5, Informative)
LeelaZero - an open source go bot, has beat 9d professionals and other lower ranked professionals. It is also ranked #3 in the world in gobot competitions, and that was with using half or less of the hardware resources that many of hte competitors had (LeelaZero was using 4 1080 TI GPUs; the competitors had 10 1080 TI GPUs).
It still hasn't reached the level of AlphaZero, but if you'd like to help it do so, you can contribute here.
http://zero.sjeng.org/ [sjeng.org]
Note that they benchmarked against LeelaZero, but had it misconfigured - they gave their bot 80,000 playouts, and LeelaZero 50 seconds per move, but left a default where LeelaZero doesn't use all of its time. So often it was moving in 3 seconds. It might well be weaker than LeelaZero on similar hardware when LeelaZero is correctly configured.
Re: (Score:2)
They are going to rerun the competition vs LeelaZero with the recommendations made by LeelaZero coders (the tuning algorithm that will set it to run max speed on the hardware, and the command line settings that will ensure it runs for the full 50 seconds per move).
From the discussions on the github, it sounds like it should still beat Leela by a significant amount (predictions are 90% of games) though perhaps not as badly as their initial run (100% of games).
Online Go tournaments? (Score:2)
Since they intend to go Open Source with this, does that effectively mean that anyone with access to sufficient computing power could then use this to cheat at online Go tournaments for money? Are there such things?
PS: Asking for a friend...
Re:Online Go tournaments? (Score:4, Interesting)
People are already using LeelaZero (the #3 ranked in gobot competition and capable of beating 9d professionals) to cheat, but when a human plays drastically better than their rating, the cheating is pretty obvious.
Re: (Score:2)
People are already using LeelaZero to cheat
But do they win money? I have never seen an on-line tournament with money on the table.
but when a human plays drastically better than their rating, the cheating is pretty obvious.
I like to play in parallel with the computer. I decide on my best move, then I look at the move the computer selected, and try to understand why it is better. This technique has improved my game a lot. I tend to focus on responding to my opponent's move, while the computer looks at the entire board and will often make an unrelated move on the other side of the board. I have tried to focus on doing the same. Often th
Re: (Score:1)
Re: (Score:1)
Re: (Score:2)
but when a human plays drastically better than their rating, the cheating is pretty obvious.
So the next task should be to create a bot where you can set a target rating that it will play at. Then you can slowly move up the ranks to simulate your play improving over time.
Dang (Score:2)
Gobots suck (Score:3)
Re: (Score:2)
Damn, I was about to post that.
+1 First Post to you.
Good question (Score:3)
Obviously the answer is "nothing" since they just created another game playing program. What a joke.
Re: (Score:1)
""But we wondered: Are there some unanswered questions? What else can you apply these tools to?"
Obviously the answer is "nothing" since they just created another game playing program. What a joke.
And made the source code available to others to adapt as they see fit. We haven't found out yet what others will apply this to, but I think it's too early to conclude it'll be "nothing".
Re: (Score:2)
We haven't found out yet what others will apply this to, but I think it's too early to conclude it'll be "nothing".
Turn-based positional board games are all of a kind, as far as machine learning is concerned.
OTOH, that StarCraft bot mentioned in TFS is a different kind of problem solving. It would be quite interesting to see how a machine learning-based bot stands up against the in-house bots written by professional players.
Solving a solved problem (Score:1)
Solving a problem that has already been solved is not as difficult as solving a problem that has not been solved.
I understand that they don't have skills to do anything new, but why do they bother doing this? Google has a goal of generating general AI, which is why they are doing this. Facebook is just copying them.
Some info on leela-zero vs. ELF (Score:2)
Starcraft? (Score:2)
Facebook's AI Research group is also developing a StarCraft bot that it too plans to open source.
What would be the point of this? I'm not any sort of expert in starcraft but I'm not sure what this would prove. While there is a lot of tactics and strategy to work out, a huge part of a game like this is simply the ability to click on and order units about as fast as possible. A computer could very obviously do this faster than any human unless it was artificially limited. I have a hard time seeing how a human could keep up with a computer speed zerg rush. Is there something I don't understand about
Re: (Score:2)
There are quite a few SC bots written in-house by professional players (well, mostly former professionals, who reach their elderly late-20s and aren't fast enough any more). It would be very interesting to see how a machine learning-based bot competes against bots coded with specific human-like strategies. Both would have unlimited APM, so no advantage there, but how would the machine learning bot play? I'd imagine micro would be similar, as there's apparently one best way to do that, but macro? That's
Re: (Score:2)
> While there is a lot of tactics and strategy to work out, a huge part of a game like this is simply the ability to click on and order units about as fast as possible
= Short Answer =
TL:DR; False. Even 19,000 APM [technologyreview.com] won't save you.
= Long Answer =
First off, a few terms [liquipedia.net] so those that are't familiar with StarCraft aren't completely lost:
* APM = Acronym for Actions Per Minute. How fast you can click.
* micro -- ability to control your units individually (i.e. tactical positioning of units taking advantage of how
Go was super hard! (Score:2)