if you understand the memorization I mentioned, then you know the answers

送交者: conner 于 2016-03-16, 08:47:45:

回答: 这和我理解还是一致的。不过有些问题，不知各位专家可否解答、讨论由 pgss 于 2016-03-16, 00:41:30:

every move Lee made will trigger the updating of the searching tree. Tree branch not relevant any more is trimmed, existing branches's Bayesian values are updated, and the tree grows wide and deep. I assume the tree is also growing wide and deep when Lee is thinking about the next move. If Lee's move keeps stay in a main branch of the tree then the tree gets deeper and stronger. If not, then the tree gets weaker. If Lee's move is not in any branch, then the tree needs to be rebuilt.

That was why I said AlphaGo is playing with/against itself. And Lee also felt that he played against himself.

所有跟贴:

奇怪的事情就在这里 - 008 (252 bytes) 2016-03-16, 09:06:51 (804024)
- 真有可能是狗在作弊放水。那我就输得冤了。 (无内容) - pgss (0 bytes) 2016-03-16, 10:24:16 (804033)
  - no chance (无内容) - conner (0 bytes) 2016-03-16, 10:27:05 (804036)
- 唯一解释就是它认为这是败招，不需建树。 (无内容) - pgss (0 bytes) 2016-03-16, 09:38:36 (804026)
  - no, only when the tree grow decently deeper from there - conner (161 bytes) 2016-03-16, 09:46:17 (804027)
    - but it wasn't in the tree. hassabis said. (无内容) - pgss (0 bytes) 2016-03-16, 09:59:40 (804030)
      - what is the it? the move 78? (无内容) - conner (0 bytes) 2016-03-16, 10:07:05 (804031)
        
        Yes。he said 78 is less than one in 10k chance in its estimate (无内容) - pgss (0 bytes) 2016-03-16, 10:15:49 (804032)
        
        then what is the criteria to build a new scenario? (无内容) - pgss (0 bytes) 2016-03-16, 20:44:19 (804062)
        
        That means it was in the tree with a lower Bayesian value - conner (89 bytes) 2016-03-16, 10:26:11 (804034)
- the program has a time limitation (run limitation) for each move - conner (366 bytes) 2016-03-16, 09:14:40 (804025)
  - The time limit was 5 seconds per move when against Fan Hui (无内容) - conner (0 bytes) 2016-03-16, 12:47:05 (804048)
  - 当对手下出意料之外的棋 - 008 (442 bytes) 2016-03-16, 11:49:15 (804043)
    - time/resource allocation during a game is a deep-deep problem - conner (377 bytes) 2016-03-16, 12:05:27 (804046)
      - It is like to predict how long 'Lee' and itself will take for next N moves (无内容) - conner (0 bytes) 2016-03-16, 12:10:12 (804047)
  - make no sense again (无内容) - 008 (0 bytes) 2016-03-16, 10:33:37 (804037)
    - the time limit for each move or tree operation? (无内容) - conner (0 bytes) 2016-03-16, 11:50:12 (804044)
      - time limit (无内容) - 008 (0 bytes) 2016-03-16, 12:00:06 (804045)
thanks. but after a move isnt it already a new scenario? (无内容) - pgss (0 bytes) 2016-03-16, 08:50:18 (804022)
- No, it is very likely still in the tree. The tree is memorized and updated. (无内容) - conner (0 bytes) 2016-03-16, 08:52:05 (804023)

加跟贴