AlphaGo - Lee Sedol 4:1

[Robert Jasiek]

The final result of the match AlphaGo - Lee Sedol is 4:1. The computer
program beats the top human player.

[sobriquet]

So we've finally entered the era of 9d bots.

It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

[Robert Jasiek]

Post by sobriquet
It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

It needed more than two wise ideas.
- MC
- using NN at all
- calibrating NN
- learning from strong games
- creating initial values
- reinforcement learning by self-play
- sampling only one move per sample game
- iterating self-play and sampling
- combination of everything
- probably more

[sobriquet]

Post by Robert Jasiek

Post by sobriquet
It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

It needed more than two wise ideas.
- MC
- using NN at all
- calibrating NN
- learning from strong games
- creating initial values
- reinforcement learning by self-play
- sampling only one move per sample game
- iterating self-play and sampling
- combination of everything
- probably more

Yeah, that just sums up some activities that boil down to 'combining neural networks with MCTS in a smart way'.

[Bill]

Post by sobriquet

Post by Robert Jasiek

Post by sobriquet
It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

It needed more than two wise ideas.
- MC
- using NN at all
- calibrating NN
- learning from strong games
- creating initial values
- reinforcement learning by self-play
- sampling only one move per sample game
- iterating self-play and sampling
- combination of everything
- probably more

Yeah, that just sums up some activities that boil down to 'combining neural networks with MCTS in a smart way'.

Neural networks are discrete as are some of the related processes.
Truly capturing (no pun intended) the whole of the gamespace seems like
a remote possibility.
We can probably do better and better and better but, increase the size
of the board and we where we stand. I applaud for what HAS been
demonstrated.

Bill (16K)

[Hans-Georg Michna]

Post by sobriquet
So we've finally entered the era of 9d bots.
It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

9p, not 9d.

Hans-Georg

[John Geoffrey]

Post by Hans-Georg Michna

Post by sobriquet
So we've finally entered the era of 9d bots.
It looks like it took two revolutions to get the bot skills from 1k
to 9d. First with the introduction of MCTS, that allowed bots to go
from about 1k to 5d and the next revolution was the integration of
neural networks and combining these with the MCTS algorithm in a
smart way and this allowed bots to advance from 5d level to 9d level
and beyond.

9p, not 9d.
Hans-Georg

yes, was it mentioned here already? AlphaGo received an honorary 9d
title from the Korean Go association.

[Robert Jasiek]

AlphaGo received an honorary 9d title from the Korean Go association.

The KBA is open-minded to new developments. In this case, a 9p for a
"player" with funny-looking kyu-like moves during the endgame when
ahead, play becoming desparate when greatly behind instead of knowing
to resign (but Lee did likewise in game 3) and "no manners" for the
first move (ok, this is really immaterial) :) The program does not
know good timing of privileges, of advanced aji exploitation and
probably the meaning of at least two exchanged kos on the board yet.
It "only" knows how to win with Black or White;)