Gomoto wrote:
When I play I dont care about influence. My friends care about influence. I care about good moves. I win.
![Wink ;-)](./images/smilies/icon_e_wink.gif)
LZ doesn't care about influence either... So she isn't one of Gomoto's friends?? i thought everybody loved LZ!
The human brain learns by forming new synapses, and modifying existing ones, a process not dissimilar to the way a DCNN modifies its weights.
LZ figures out how to modify its DCNN weights by reading all the way to the end of the game, seeing who won, and backing up that information in the context of all the other variations in the forest it has looked at.
Even if Gomoto doesn't read all the way to the end of the game at every move, he at least later on finds out whether he won or lost in actual play (a forest of width = 1).
So if he learns anything from his wins and losses, Gomoto could be said to learn Go in the same kind of way that LZ does, because (a) they both learn by modifying the strengths of synapses, and (b) they both don't care about influence.
https://github.com/gcp/leela-zero/blob/ ... #compiling says:
"Recomputing the AlphaGo Zero weights will take about 1700 years on commodity hardware."
Let F be the width of the forest of variations LZ explores.
Batman always wins - So, because he always wins against his friends, Gomoto
![batman :batman:](./images/smilies/icon_batman.gif)
must be (1700 x F) years old, or he has a brain (1700 x F) times bigger than mine.
Wait a minute!
![batman :batman:](./images/smilies/icon_batman.gif)
only cares about good moves, and we all know that the only good moves are the ones LZ makes, so
![batman :batman:](./images/smilies/icon_batman.gif)
is LZ!!!!!!
gennan wrote:
Those influence maps.. they look too crude to be useful for anything.
How does it help to detect who is leading or which areas are big and / or urgent?
An Example of Influence Map Utility: Visualising Tradeoffs between Obedient Gote and Wishful Tenuki here are two rows of pictures of different white responses to black L4, to show the tradeoff between:
top row = blocking in gote at M4 to save a couple of stones and prepare the way for a later invasion/reduction of black's bottom right corner
bottom row = taking sente to gain influence elsewhere, eg D10
Attachment:
leela226.png [ 457.37 KiB | Viewed 9481 times ]
Bill Spight wrote:
I do think that an influence map produced by having a top bot play itself from the current position 1,000,000 times could be quite educational, however
The Leela maps in the above picture are her Moyo markup. Her maps after thinking are a bit different from those before thinking. For M4, Leela changes her mind a little bit after looking at 2.5 million nodes... she made my cpu get hot, so for the second run (D10) i took a snapshot after only 700,000 nodes - short of a million, but the map was pretty stable by then; you can see it is more than a bit different from her first guess (which has black's lower group smothered by white's massive moyo!).
i like the look of what Influencie produces, as it corresponds with my own intuition - although this example might be a touch optimistic about connection strength of long jumps? Influencie lets you play around with Stone influence power; i may have had it turned up too high.
the 3 different mappers agree on a lot of things, but there are a few differences.
Influencie wins the beauty competition by a mile, and the size of her left black shadow agrees with the guess i made during my game review with W, so i am biased like hell in her favour (Oh, i remember now, i twiddled with her Stone Influence control until she did!
![Embarrassed :oops:](./images/smilies/icon_redface.gif)
) - but the other two may have sounder judgement.