w********2
发帖数: 632 | |
d*****u
发帖数: 17243 | 2 现在deep learning没啥直接的建模。
对数据的统计分布等等不关心，就是直接上神经网络，反正理论上什么函数都能拟合。 |
w********2
发帖数: 632 | 3 李飞飞没那个图库就啥也干不了，所以还是supervised
有归类，算相似分数 |
w********2
发帖数: 632 | 4 有的，在底层。建几百上千，甚至更多，然后选最符合training set的。
【在 d*****u 的大作中提到】
: 现在deep learning没啥直接的建模。
: 对数据的统计分布等等不关心，就是直接上神经网络，反正理论上什么函数都能拟合。
|
d*****u
发帖数: 17243 | 5 除了网络结构没有什么模型。deep learning的模型基本都是相似的，最多试少数几个
网络架构
然后就是调整hyperparameter
其他都靠training
【在 w********2 的大作中提到】
: 有的，在底层。建几百上千，甚至更多，然后选最符合training set的。
|
w********2
发帖数: 632 | 6 是这样的，就是调参数，但参数优化靠training set，容易overfitting。
ai现在火的关键是有了大的库，读写都很快，实用，其实相似技术1990年代就有了。
【在 d*****u 的大作中提到】
: 除了网络结构没有什么模型。deep learning的模型基本都是相似的，最多试少数几个
: 网络架构
: 然后就是调整hyperparameter
: 其他都靠training
|
w********2
发帖数: 632 | 7 举个例子，nmr样本谱图和标准库对比得分排序，一个道理。
【在 w********2 的大作中提到】
: 是这样的，就是调参数，但参数优化靠training set，容易overfitting。
: ai现在火的关键是有了大的库，读写都很快，实用，其实相似技术1990年代就有了。
|
w********2
发帖数: 632 | 8 Go program, based on a combina- tion of deep neural networks and tree search
http://web.iitd.ac.in/~sumeet/Silver16.pdf |
w********2
发帖数: 632 | 9 We have introduced a new search algorithm that successfully combines neural
network evalu- ations with Monte Carlo rollouts. Our program AlphaGo
integrates these components together, at scale, in a high-performance tree
search engine.
a new search algorithm: this is the key |
w********2
发帖数: 632 | 10 During the match against Fan Hui, AlphaGo evaluated thousands of times fewer
positions than Deep Blue did in its chess match against Kasparov4;
compensating by selecting those positions more intelli- gently, using the
policy network, and evaluating them more precisely, using the value network
—an approach that is perhaps closer to how humans play. |
|
|
w********2
发帖数: 632 | 11 Go is exemplary in many ways of the difficulties faced by artificial
intelligence33,34: a challenging decision-making task, an intractable search
space, and an optimal solution so complex it appears infeasible to directly
approximate using a policy or value function. |
w********2
发帖数: 632 | 12 By combining tree search with policy and value networks, AlphaGo has finally
reached a professional level in Go |
w********2
发帖数: 632 | 13 an intractable search space: cs
a challenging decision-making task: tree, statistics
a policy or value function: statistics |
w********2
发帖数: 632 | 14 All games of perfect information have an optimal value function, v*(s),
which determines the outcome of the game, from every board position or state
s, under perfect play by all players. These games may be solved by
recursively computing the optimal value function in a search tree containing
approximately bd possible sequences of moves, where b is the game’s
breadth (number of legal moves per position) and d is its depth (game length
).
this is brutal force playing the game rule, basically the game is broken. |
w********2
发帖数: 632 | 15 i think ml is more mathematical optimization than statistics or cs. |
w********2
发帖数: 632 | 16 In large games, such as chess (b ≈ 35, d ≈ 80)1 and especially Go (b ≈
250, d ≈ 150)1, exhaustive search is infeasible2,3, but the effective
search space can be reduced by two general principles. First, the depth of
the search may be reduced by position evaluation: truncating the search tree
at state s and replacing the subtree below s by an approximate value
function v(s) ≈ v*(s) that predicts the outcome from state s. This approach
has led to superhuman performance in chess4, checkers5 and othello6, but it
was believed to be intractable in Go due to the complexity of the game7.
Second, the breadth of the search may be reduced by sampling actions from a
policy p(a|s) that is a prob- ability distribution over possible moves a in
position s. For example, Monte Carlo rollouts8 search to maximum depth
without branching at all, by sampling long sequences of actions for both
players from a policy p. Averaging over such rollouts can provide an
effective position evaluation, achieving superhuman performance in
backgammon8 and Scrabble9, and weak amateur level play in Go10. |
w********2
发帖数: 632 | 17 In other words, in the future if computer speed increases 1 million times,
alpha go can go straight brutal force exhaustive search without any
simplification, and will
be more accurate. |
w********2
发帖数: 632 | 18 so the only way to beat the alpha go is to break the game rule by human
player, so that the alpha go has to recalculate each time. |
w********2
发帖数: 632 | 19 don't be a perfect player when play with alpha go. that is the key.
state
containing
length
【在 w********2 的大作中提到】
: All games of perfect information have an optimal value function, v*(s),
: which determines the outcome of the game, from every board position or state
: s, under perfect play by all players. These games may be solved by
: recursively computing the optimal value function in a search tree containing
: approximately bd possible sequences of moves, where b is the game’s
: breadth (number of legal moves per position) and d is its depth (game length
: ).
: this is brutal force playing the game rule, basically the game is broken.
|
w********2
发帖数: 632 | 20 In other words, try to beat the algorithm by playing oddly, than try to win
the game. And you will deadlock the machine (very long time calculating each
step), then win.
【在 w********2 的大作中提到】
: don't be a perfect player when play with alpha go. that is the key.
:
: state
: containing
: length
|
w********2
发帖数: 632 | 21 More precisely, break its assumptions.
win
each
【在 w********2 的大作中提到】
: In other words, try to beat the algorithm by playing oddly, than try to win
: the game. And you will deadlock the machine (very long time calculating each
: step), then win.
|
