Slashdot Mirror
Alpha Go Takes the Match, 3-0 (i-programmer.info)
mikejuk writes: Google's AlphaGo has won the Deep Mind Challenge, by winning the third match in a row of five against the 18-time world champion Lee Se-dol. AlphaGo is now the number three Go player in the world and this is an event that will be remembered for a long time. Most AI experts thought that it would take decades to achieve but now we know that we have been on the right track since the 1980s or earlier. AlphaGo makes use of nothing dramatically new — it learned to play Go using a deep neural network and reinforcement learning, both developments on classical AI techniques. We know now that we don't need any big new breakthroughs to get to true AI. The results of the final two games are going to be interesting but as far as AI is concerned the match really is all over.
Although in a sense it is "nothing new" (neural networks and monte carlo statistical techniques), the combination is one of the most convincing demonstrations of short-circuiting huge branching factors to arrive at what a human player would call "intuition".
Chess has a branching factor of about 35. This is small enough that if you prune the most dismal lines, you can brute force the rest to many ply and arrive at a very good result, but this is not a well generalizable technique.
Go has a branching factor of about 250. This cannot be brute forced, even with aggressive pruning. The result of a NN evaluator function plus Monte Carlo has been astonishing: it was not predicted for computer Go to reach this strength for decades yet, but here we are.
The implications of this combination of techniques to other kinds of problems requiring "intuition" will be interesting to watch.
They've already stated that their next goal is to do it again without the human database, but rather through iteration. And the big final advancement *this time* was made by it playing a ton of games with itself. Studying the human games was not enough to get to this level.
And humans do the same thing. They spend their lives studying the important games that came before. So the point is it did it pretty much the same way humans do. And it has already played a move that no strong human has ever played (Game 2, move 37). At first it (not surprisingly) appeared to be blunder, until its strength became clear. Humans will now learn from the computer and their level of play will rise. It happened in chess and checkers, and in a very big way in backgammon. Any strong human backgammon player today would trounce the World Campion of 20 years ago.
We know now that we don't need any big new breakthroughs to get to true AI.
Grossly exaggerated claim. The following article worth reading on this subject by no one else than two of authorities in the field, one did work on the backgammon game in the 90s and the other one on the IBM Deep Blue program that win over the world chess champion Garry Kasparov in 1997. http://www.ibm.com/blogs/think... In particular:
"However, research in such “clean” game domains didn’t really address most real-life tasks that have a “messy” nature. By “messy,” we mean that, unlike board games, it may be infeasible to write down an exact specification of what happens when actions are taken, or indeed what exactly is the objective of the task. Real-world tasks typically pose additional challenges, such as ambiguous, hidden or missing data, and “non-stationarity,” meaning that the task can change unexpectedly over time. Moreover, they generally require human-level cognitive faculties, such as fluency in natural languages, common-sense reasoning and knowledge understanding, and developing a theory of the motives and thought processes of other humans."
Achille Talon
Hop!
There's a well known phenomenon where every time some AI research produces a successful result, someone comes along and says "That's not true A.I" "It's just a computer program that has to be told what to do."
(This is the "no true Scotsman" argument.)
So let's see the list of such "non-AI" technologies:
- Natural-Language translation (getting pretty usable now)
- Speech recognition combined with ability to answer fairly arbitrary questions quite well on average.
(talking to Google via my Android phone)
- Self-driving cars (getting pretty close - will be better drivers than people on average pretty soon)
- Chess
- Jeopardy
- Go
- Detection of suspicious speech and patterns of communication (no doubt used by NSA on most Internet and phone traffic)
- Recognition of particular writer from all writers on Internet by analysis of their writing style
- Person identification by face picture recognition
- Object type and locaton type recognition from pictures
- Walking, box-stacking robot "Atlas 2"
Just algorithms.
Does it actually matter what you personally choose to call this kind of technology? It is what it is, and it's advancing quickly.
"It's not true AI" sounds like the desperate retreat cry of a person in a very defensive stance, afraid of losing a sense of human uniqueness.
Where are we going and why are we in a handbasket?
Well, yes and no. Back when Deep Blue beat Kasparov in 1997 it was programmed with a huge amount of chess logic programmed by people. Using a computer amplified the power of those algorithms, it had move databases but it wasn't really self-modifying. From what I understand you could step through the algorithm and even though you couldn't do it at the speed of the computer, you could follow it. That approach pretty much failed for Go, it's very hard for a human to quantify exactly what constitutes a good or bad move.
Neural networks pretty much does away with that in any form humans can follow. That is to say, if you had to explain how Alpha Go plays you'd get a ton of weights that don't really make much sense to anybody. It means you don't need Go expertise in the programming, because they couldn't find where to tweak a weakness even if they saw one. All you can really do is play it and it'll learn and adjust from its losses. From what I've gathered it's hard to find excellence, if you train with lots of mediocre players making mediocre moves it's easy to learn decent moves but that'll fail against a master. And that if you let it self-play it can easily learn nonsense that'll only work against itself.
Apparently they've solved those problems well and has now created a machine that plays at a beyond-human level. If they can extend this approach to practically unlimited choices like say an RTS where you can choose what to build, where to send your units, when to attack, when to defend, what resources to collect etc. it could be absolutely massive. Imagine if you were in say city planning and you have tons of data on traffic patterns, congestion and how traffic reflows when you open and close roads and you could put an AI on the job to say where and how you get the most value for money. I'm not sure if it's strong AI, but it's certainly places we use HI today.
Live today, because you never know what tomorrow brings
I don't count any of those things as AI, although they are components of AI since they are all pieces of (or combinations of) observing and interacting with the environment.
To me, "true AI" is something that can decide to do something other than that for which it was constructed. Can AlphaGo do anything other than play Go? If you tell it to play Go, can it decide, "No thanks, I'd rather cure cancer, it's a more rewarding problem"?
While AlphaGo and the like are very fantastic achievements, I don't think they are intelligence - they are "merely" very effective specialized problem-solving machines.
"There are a dozen opinions on a matter until you know the truth. Then there is only one." - CS Lewis (paraprhase)
So far, every post I have read that makes the same claim as yours lacks a critical piece: a clear description of what would qualify as AI.
Often, when I state that question, I get a long, rambling, disorganized list of random things humans do, and no indicator that making a computer do them would yet qualify as true AI. That is why I keep emphasizing the word "clear." Make it clear or you are just being religious.
So, exactly where are those goal-posts?