American Go E-Journal » Master/AlphaGo Commentaries

AlphaGo Zero, the latest version of the go-playing AlphaGo AI, defeated all previous versions of AlphaGo in just forty days. More importantly, AlphaGo Zero was taught go’s rules, but given no additional instructions, instead learning the best moves by playing millions of games against itself. Details of the new program were published Wednesday in the journal Nature.

“AlphaGo Zero independently found, used and occasionally transcended many established sequences of moves used by human players,” write the AGA’s Andy Okun and Andrew Jackson in an accompanying article. “In particular, the AI’s opening choices and end-game methods have converged on ours — seeing it arrive at our sequences from first principles suggests that we haven’t been on entirely the wrong track. On the other hand, some of its middle-game judgements are truly mysterious and give observing human players the feeling that they are seeing a strong human play, rather than watching a computer calculate.”

The ability to self-train without human input is a crucial step towards the dream of creating a general AI that can tackle any task, reports Nature. In the nearer-term, though, it could enable programs to take on scientific challenges such as protein folding or materials research, said DeepMind chief executive Demis Hassabis at a press briefing. “We’re quite excited because we think this is now good enough to make some real progress on some real problems.”

Prof Satinder Singh, a computer scientist at Michigan University, who reviewed the findings for the journal said: “The AI massively outperforms the already superhuman AlphaGo and, in my view, is one of the biggest advances, in terms of applications, for the field of reinforcement learning so far.”
– Chris Garlock; Image credit: Saran_Poroong Getty Images

Read more…

The AI That Has Nothing to Learn From Humans
DeepMind’s Go-playing AI doesn’t need human help to beat us anymore
AlphaGo Zero: Google DeepMind supercomputer learns 3,000 years of human knowledge in 40 days
DeepMind has a bigger plan for its newest Go-playing AI
AI versus AI: Self-Taught AlphaGo Zero Vanquishes Its Predecessor
AlphaGo Zero Goes From Rank Beginner to Grandmaster in Three Days—Without Any Help
‘It’s able to create knowledge itself’: Google unveils AI that learns on its own
DeepMind AlphaGo Zero learns on its own without meatbag intervention
This more powerful version of AlphaGo learns on its own

“Apparently, some people believe that someday a human will be able to defeat AlphaGo,” writes Joel Sanet. “It’s not gonna happen. The reason is biological, not technological. No human being is capable of thinking about the game the way AlphaGo does. AlphaGo’s way of thinking is better than the human way; ergo it is no longer possible for a human to beat AlphaGo. We human beings are not capable of considering a choice of moves by determining a concrete number for each called “the probability of winning” then choosing the one with the highest value, but this is what AlphaGo does.

“Thinking that it is possible for a human to win now is due to anthropomorphization, the application of human attributes to something that is not human, a process rampant in the go community. I have heard people say, ‘AlphaGo likes the early 3-3 invasion’ or ‘He (or she) likes thickness.’ AlphaGo can’t ‘like’ anything because it has no emotions. It plays the early 3-3 invasion because it maximizes its probability of winning in certain openings. Also, as far as I know, AlphaGo has no concept of thickness. It has nothing to do with how AlphaGo derives its moves. Furthermore, AlphaGo is not a ‘he’ or a ‘she’. AlphaGo is an ‘it’.

To attribute thinking to AlphaGo is also a mistake. I wrote that it chooses the option with the highest probability of winning. It doesn’t “choose” anything because it isn’t self-aware. AlphaGo receives input, does what it is programmed to do, and produces output. To me this is more akin to a human knee jerk than to true thought. A doctor’s percussion hammer causes sensory neurons to fire off a signal to the spinal cord where it is processed and returned to the knee via motor neurons without intercession of the brain. This is analogous to AlphaGo’s input-programming-output. AlphaGo’s programming is immutable. The day AlphaGo changes its own programming is the day I’ll say it thinks.

Nevertheless, humans can learn from AlphaGo. We have learned that the shoulder hit is a lot more useful than anyone thought. AlphaGo’s new 3-3 invasion joseki makes sense so we can benefit from that, but I advise you not to do the early invasions until you are able to read the rest of the game to the end.

Alphago’s supremacy over humans is no reason to feel that studying go is a dead end. Your study is de facto open-ended because you will never reach the end of it. People study go to improve, not to become the strongest player on the planet.”