送交者: 短江学者 于 2016-03-15, 22:10:41:

回答: 蓝海你真敢跟我讨论人工智能Montecarlo的事？ 知道天高地厚吗？ 由 短江学者 于 2016-03-15, 21:53:10:

• 另外这是我的前贴最后一句说了mc - 短江学者 (542 bytes) 2016-03-15, 22:13:28 (804002)
  • 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 - pgss (731 bytes) 2016-03-16, 00:41:30 (804013)
    • if you understand the memorization I mentioned, then you know the answers - conner (602 bytes) 2016-03-16, 08:47:45 (804020)
      • 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)
  • sorry 自然文章方程没有号将就看看abstract吧。 - 短江学者 (1508 bytes) 2016-03-15, 22:24:05 (804003)
    • if reinforcement learning is an inherent physioneurological mechanism - cornbug (45 bytes) 2016-03-16, 01:28:11 (804016)
      • it is because their craziness have not be properly evaluated (无内容) - conner (0 bytes) 2016-03-16, 08:49:19 (804021)
    • 自然文章标题和作者 - 短江学者 (485 bytes) 2016-03-15, 22:55:43 (804005)
    • 如何使用 policies和如何学习 policies 是两回事情。 (无内容) - bluesea (0 bytes) 2016-03-15, 22:55:16 (804004)
      • 懂不懂自然文章也是两回事情。增强学习主要思想就是边用边学。 (无内容) - 短江学者 (0 bytes) 2016-03-15, 22:59:16 (804007)
        • 所以你认为，增强学习是alphago的几乎唯一或者最基础和主要的算法框架。 - bluesea (20 bytes) 2016-03-15, 23:08:41 (804009)
          • 应该是吧。棋谱和初始的policy network只是收缩范围，减少选项 (无内容) - pgss (0 bytes) 2016-03-15, 23:21:27 (804012)
          • 是他们文章标题和摘要说的。我只能看文章说话。 (无内容) - 短江学者 (0 bytes) 2016-03-15, 23:15:10 (804011)

加跟贴