Tuesday assorted links


The magical nature of The Beatles' vocal harmonies tends to obscure the brilliance of the bass & guitars . . .

#2 Environmental Science waif quote to wrap up the article:
"“If money is involved, some people will act dirty anyway,” Nina Bakker, a student of environmental issues at the University of Amsterdam, said as she paused during a bike ride in the center of the city. “With all the scandals, why should they all of a sudden start acting like decent people?”"

What a sweet, dear, precious little thing she is.

Is it too early to hope she ends up like those Party Loyalists in Solzhenitsyn's work? Laboring for 18 hours a day in freezing water while proudly proclaiming their loyalty to the socialist system?

5. I'm resisting the urge to say "1985 called, and it wants its trendy technique back," because we really do have expanded abilities to work with neural nets these days.

To answer your question, probably not without tree search, but maybe as a component of tree search. They suggest as much in their conclusion, saying it might be useful for move ordering. At this point it's unresolved. Serious game people are not dismissing it out of hand.

Training it with pro games then testing it against mid-kyu level amateur programs might not be a really fair test.

1.) Amateur -kyu and low -dan games really are a completely different game than pro level games. To the point where pro level games don't even make sense to amateurs until things come together right at the end.

2.) All the previous Go algorithms that competed against pros get most of their advantage through non-sequitor moves. Their moves weren't unexpected, but effective shifts in focus (tesuji), they were just weird and threw the pros off their game in the one-off games they played. If you show up in a neon pink suit with a periwinkle mohawk, chewing bubblegum and listening to Slayer on headphones, you might beat a pro in a casual game just because he's weirded out. That doesn't mean you are better, or the pro wouldn't be able to get himself together and wipe the floor with you in a series of games. Psychology has an impact, especially in the short run.

That being said, I'm really interested in seeing how effective this system is against players at the level it studied, or how it does against computers if it's trained on computer Go games instead of human pro games.

As to the question in your last paragraph, I'm sure it would get thrashed. It's badly beaten by Fuego right now. Any significant handicap is going to castrate it.

But MCTS, and really any tree search, is potentially helped by a good move generator. Which this or something like it might turn out to be.

And frankly, regardless of the underlying techniques, the smart bet over the long term is on the machine. And the long term in this case is 50 years on the outside and probably realistically within the next 15 years.

I agree that GO needs to go the way of computer chess. I think this hybrid alpha-beta tree plus pattern recognition approach highlighted by the article is interesting. And it's high time arrogant GO players, who think they are doing something spiritual, get beaten by the machines. BTW chess, which relies on calculation rather than pattern recognition, is IMO superior to GO but let's not start a flame war.

Pattern recognition is a calculation.

"Should they rid New Zealand of non-human mammalian predators?" How in heaven's name could they achieve that?

"New Zealand's Manhattan Project"

Apparently by dumping poison everywhere.

Gotta burn down the village to save it, I guess.

When they get serious, they will start releasing genetically altered diseases that render the females sterile.

A bit tough on the bats though.

Didn't Australia do something similar with rabbits. Ended up with a new population that was resistant to the disease.

Myxomatosis wiped out most of them in the '50s. That was a life-saver - the numbers were disastrously high at that point and the virus wiped out over 80% of the population. It took the rabbits 40 years to bounce back to half their previous population high. Most are resistant at this point, but they're hitting them with new biological weapons these days - calcivirus, GM-viruses to induce infertility, etc. Same for cane toads, mice, carp, foxes, etc.

"How in heaven’s name could they achieve that?"

Chinese needle snakes of course.

Won't work in New Zealand, the winters aren't cold enough to kill off the gorillas we're going to need to eat the snakes.

The neural network go player can't even beat the best tree search player more than 10% of the time. And the best tree search player isn't even 1 kyu -- which means that it only has a 50% chance of winning against the best human players (9 dan) if they take a 9-stone handicap, which is huge. So no, we're not really that close.

It depends on what you mean by close. Progress is about one stone per year, so we're about a decade away from the Deep Blue moment for Go.

But to your point, if this tool is important it will be as a part of some tree search program which is at least close to asymptotically perfect with increasing resources.

I don't really care if a computer can beat anyone at go. I care if a computer can accurately translate between languages, or autonomously clean my house without doing anything crazy, or drive me to work without killing me. We're getting close to one of those, at least.

Hopefully you mean the cleaning one because otherwise you are in for a world of disappointment.

I'm compelled to point out that the above-linked article on computer Go is an absolute barrage of nonsense, and includes numerous inaccuracies that could have been avoided by even a few minutes of Googling. The arXiv paper that it discusses features a computer program that is significantly weaker than already-existing Go software, which themselves are nowhere close to defeating even a low-level human professional.

In general, I find the arXiv blog to be little more than a hype machine that regurgitates the dubious opinions of the authors it covers, with no fact-checking whatsoever. I wouldn't believe anything written on it without some other sources, or at least a close reading of the paper that is discussed (arXiv papers themselves are notorious for being full of BS, at least those that are not subsequently published in real journals).

Link to the neuralnet GO paper: http://arxiv.org/abs/1412.3409

Here is a link to Randy bacman explaining the opening chord. There is a better video where you can see him play it. I'm on my phone and can't find it right now. https://www.google.com/url?sa=t&source=web&rct=j&ei=86mQVPf3EcKgyAS20YK4DA&url=http://m.youtube.com/watch%3Fv%3DgwfH9oAiPH0&ved=0CCEQyCkwAA&usg=AFQjCNG9oaZpneCIORL8844yo2yRUgkm5w&sig2=wzmy4P-CAcRww6mHMRr18Q

Utterly appalling fact checking error in the New Yorker article.

"It consists of two major islands—the North Island and the South Island, which together are often referred to as the mainland—and hundreds of minor ones."

The South Island is the "Mainland" (and it's denizens "Mainlanders"). The North Island is never referred to as the mainland.

4. Indubitably. On any rainy day I see several dozen umbrellas cast aside as broken. It seems most umbrellas are single use.

The US Army is the last holdout of military forces that permit the use of umbrellas in uniform. The day I got poured on in Mess Dress uniform was the day I was convinced the taboo was ridiculous, at least for dress uuniforms. In BDUs, an umbrella was largely unnecessary as long as you wore a gortex rain jacket. Before that, you'd get pretty wet in field jacket and BDU cap or beret, but you were a soldier and just didn't care about being wet. It was part of the job.

#4 The umbrella that cannot be bent. Pity the surrounding crowd. Should have coupled it with a light sabre.


Indeed, what happens when all the people around you are using air umbrellas and deflecting their rain onto you? Answer: get Air Umbrella-D, a behemoth which outblows their puny umbrellas. And thus an umbrella arms race is precipitated.

Hmm. Maybe someone will offer an end cap converter for those infernal leaf blowers.

#1 To my untrained ear, the chord he plays is not quite the same as the original, and not as rich and interesting. So, in my opinion, the mystery remains unsolved.

I agree, especially when I listen to Randy Bachman's re-creation of the chord, which sounds much closer to the original than the mathematician's does.

#6 Despite a few minor errors a good article

Assuming the majority of the readership of MR is North American I would point out that we have no season on hunting deer here. Lots of reds, plenty of sika and even the odd wallaby.
Firearms are licensed and pistols are much rarer than US but plenty of hunting for those who are keen

#6 -seems like New Zealanders, like their Aussie cousins, are into violence and killing things (AUS is #1 for OECD violence rates). But the conservationist in the NY article is not going to dent unwanted species much in his country, though he does make his little farm a sort of sanctuary for indigenous fauna. I for one am for exterminating rabbits in Australia, invasion species mammals in NZ, and tree snakes in Guam and the Norwegian (or is it the Swedish?) rat in Hawaii, but it will be a tough effort that frankly needs international money and cooperation that's not going to come due to PETA and other such misguided outfits.

Peta = Crazy Cat Ladies

Does Australia have the highest violence rates in the OECD? Only in one horrible category and that's rape. I suspect and hope that our reporting rates are higher than in other countries, but comparing reporting rates is obviously difficult. On the brighter side our murder rate is about 24% that of the US and our robbery rate is 14% that of the United States. We are 25% more likely to beat each up (commit assault) than the US, however. But that's probably not a surprise.

My experience (Aus, NZ and USA) bears out the statistics, that the USA is less violent in terms of street fights, brawls etc, despite having a higher homicide rate. One theory is that with the far greater availability of guns in the US, physical confrontations are more likely to turn out deadly, so there is more reticence to get into a fight.

A few corrections on the absurdly hyperbolic go article:
The rank of 1 kyu does NOT represent a professional expert. Amateur rankings go from down to 1 kyu, and then up from 1 dan to 7 dan (1d-7d). Proffesional rankings go from 1 dan to 9 dan (1p-9p), with 1p being at least no worse than 7d. A rank of 5-4 kyu does not represent many years of study; for instance I achieved this rank after about 9 months of casual play. Gnu Go and Fuego are not even close to two of the strongest programs; several modern Monte Carlo Tree Search engines play online ad 2d, and one, Crazy Stone, has an online rank of 6d. While I am excited about the use of pattern recognition based approaches in go, especially to seed starting moves for MCTS, there is still significantly more of a ways to go before it has any appreciable effects on high-performance computer go than this article would have one believe.

Comments for this post are closed