Go has fallen to computer domination?

[Laskos]

Synopsis of top Go professional's analysis of Google's Deepmind's Go AI


A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

One stone at this level is maybe worth 300 Elo points.

The 2 hour analysis itself is on YouTube:

But I have difficulty understanding both his English and his explanations.

[Uri Blass]

Synopsis of top Go professional's analysis of Google's Deepmind's Go AI


A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

One stone at this level is maybe worth 300 Elo points.

The 2 hour analysis itself is on YouTube:

But I have difficulty understanding both his English and his explanations.

I understood
Fan Hui who lost 5-0 is 2p professional

Do they include Fan Hui in "all pro's who've looked at these games" or maybe the intention in "all pro's" is only for a significant higher level?

[towforce]

A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

How quickly is AlphaGo improving? It seems to me that it went from "ordinary computer level" to "extraordinary computer level" fairly quickly - and if the rate of improvement hasn't slowed down too much, these Go experts might be being presumptious in the same way that chess GMs used to be until the Kasparov match against Deeper Blue.

[Laskos]

Synopsis of top Go professional's analysis of Google's Deepmind's Go AI


A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

One stone at this level is maybe worth 300 Elo points.

The 2 hour analysis itself is on YouTube:

But I have difficulty understanding both his English and his explanations.

I understood
Fan Hui who lost 5-0 is 2p professional

Do they include Fan Hui in "all pro's who've looked at these games" or maybe the intention in "all pro's" is only for a significant higher level?

I think Kim (9p) refers to stronger professionals, although Fan Hui is no slouch judging from his (different from dan) Elo level. Korean professionals have more rigorous standards, besides that, a hard earned 9p professional probably talks about similar level to himself (maybe even higher).

[Laskos]

A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

How quickly is AlphaGo improving? It seems to me that it went from "ordinary computer level" to "extraordinary computer level" fairly quickly - and if the rate of improvement hasn't slowed down too much, these Go experts might be being presumptious in the same way that chess GMs used to be until the Kasparov match against Deeper Blue.

On similar hardware, AlphaGo is ~1000 Elo points stronger than old MCTS engines like Crazy Stone and Zen. On the cluster, it's another ~300 Elo points stronger. So, it's a gigantic leap, but it seems that it is still short by at least 300 Elo points from Lee Sedol. If an example would be the new at that time (2006-7) MCTS brand of strong engines, then one can expect 300 Elo points engine improvement in its first year of existence. I am not sure AlphaGo can improve by this much in time for Sedol match, but AlphaGo might use a stronger hardware too for Sedol match (although more hardware cannot bring very much).

[towforce]

A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

How quickly is AlphaGo improving? It seems to me that it went from "ordinary computer level" to "extraordinary computer level" fairly quickly - and if the rate of improvement hasn't slowed down too much, these Go experts might be being presumptious in the same way that chess GMs used to be until the Kasparov match against Deeper Blue.

On similar hardware, AlphaGo is ~1000 Elo points stronger than old MCTS engines like Crazy Stone and Zen. On the cluster, it's another ~300 Elo points stronger. So, it's a gigantic leap, but it seems that it is still short by at least 300 Elo points from Lee Sedol. If an example would be the new at that time (2006-7) MCTS brand of strong engines, then one can expect 300 Elo points engine improvement in its first year of existence. I am not sure AlphaGo can improve by this much in time for Sedol match, but AlphaGo might use a stronger hardware too for Sedol match (although more hardware cannot bring very much).

The human brain appears to provide an example of the scalability of NNs.

If they can make the NN bigger, with more learning capacity, then they might be able to train it, by playing it against different versions of itself, to evaluate Go positions more accurately.

The fact that AlphaGo gets better as the game progresses also implies to me that old-fashioned game tree search is also important, because obviously as a Go game progresses, the number of move choices declines.

[Laskos]

A small bit:
"Myungwan Kim (9p) says that with all his respect to the Google team, he thinks AlphaGo as it played against Fan Hui will have no chance against Lee Sedol. He says all pro's who've looked at these games generally agree that AlphaGo would need a one or two stone handicap against Lee Sedol."

How quickly is AlphaGo improving? It seems to me that it went from "ordinary computer level" to "extraordinary computer level" fairly quickly - and if the rate of improvement hasn't slowed down too much, these Go experts might be being presumptious in the same way that chess GMs used to be until the Kasparov match against Deeper Blue.

On similar hardware, AlphaGo is ~1000 Elo points stronger than old MCTS engines like Crazy Stone and Zen. On the cluster, it's another ~300 Elo points stronger. So, it's a gigantic leap, but it seems that it is still short by at least 300 Elo points from Lee Sedol. If an example would be the new at that time (2006-7) MCTS brand of strong engines, then one can expect 300 Elo points engine improvement in its first year of existence. I am not sure AlphaGo can improve by this much in time for Sedol match, but AlphaGo might use a stronger hardware too for Sedol match (although more hardware cannot bring very much).

The human brain appears to provide an example of the scalability of NNs.

If they can make the NN bigger, with more learning capacity, then they might be able to train it, by playing it against different versions of itself, to evaluate Go positions more accurately.

The fact that AlphaGo gets better as the game progresses also implies to me that old-fashioned game tree search is also important, because obviously as a Go game progresses, the number of move choices declines.

Well, AlphaGo combines the policy and value networks with MC rollouts tree search, therefore it comes natural it's playing better in the endgame. Both Supervised Learning (from human games) and Reinforced Learning (self-play) can be enhanced for all networks, policy, fast rollout policy, value. When learning from human games, there is a compromise between more samples and higher quality of samples, and they may improve on that. Small improvements in accuracy give large improvements in playing strength, for example an improvement in fast policy from 24% to 27% may well bring 200 Elo points. At the same time, scaling with more hardware, although better than with alpha-beta, becomes poor with clusters of hundreds of machines. They might improve on that.

[Daniel Shawul]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

[Laskos]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

That would be pretty amazing. Each major component of their system gave above 500 Elo points boost, and the difference between AlphaGo and Lee Sedol was less than 400 Elo points in October. They also can improve incrementally. AlphaGo is very loosely tuned. For example, regarding the balance between MCTS rollout and Value Network, they used 50/50. The paper did not mention optimization of this balance. They just tested between pure VN, pure rollout or 50/50. Optimum might well be 80/20 and only this tweak might bring additional 100+ Elo points. I think the team members are pretty aware of AlphaGo current strength relative to Lee Sedol, so if they are confident, it must give an indication. Also, the Elo span traversed by this thing with each new component and incremental improvements is huge. It may end up by March being either 300 Elo points stronger than Sedol or 200 Elo points weaker, and the result pretty conclusive one way or another.

[Ajedrecista]

Hello:

The date of the match is announced:

When is the Lee Sedol challenge match being held?

The 5-game challenge match will be held in Seoul, South Korea 9th-15th March. Match days: 9th, 10th, 12th, 13th, 15th March.

Less than five weeks from now!

Regards from Spain.

Ajedrecista.