ihappy (hungry christmas) 于Mon Dec 30 11:01:42 2002)
提到：

呵呵，我找了一下也沒找到，不過url事這個
http://www.cs.toronto.edu/~roweis/csc2515/notes/lec12x.ps.gz

【 在 adson (自強乃報國之本) 的大作中提到: 】
: 呵呵，多謝師兄指點。
: 
: 我把guitar發(fā)過的所有文章都翻出來了，
: 可還是找不到那個鏈接......
: 那個ppt大致意思說得是什么？
: 
: 【 在 ihappy 的大作中提到: 】


daniel (飛翔鳥) 于Mon Dec 30 12:23:29 2002)
提到：

【 在 adson (自強乃報國之本) 的大作中提到: 】
: excuse me, what is NFL?

NFL just means you get some goodness but sacrifice others. It should 
not be anticipated to get something good in every aspect or in every 
way. This idea is not new. Even in 1980s', researchers have implicitly 
regarded it as a norm. But until D. Wolpert, it has not been rigorously 
justified. So, at present NFL theorem is regarded as the achievement 
of him. It is worth noting that NFL does not conflict ML research because
NFL just states that there is no universal winner, which is obvious
correct. But for specific scenarios, there may exist a winner.

for NFL, refer: 
D.H. Wolpert and W.G. Macready. No free lunch theorems for optimization. 
IEEE TEC97, 1(1)
M. Koppen, D.H. Wolpert, and W.G. Macready. Remarks on a recent paper
on the no free lunch theorems. IEEE TEC01, 5(3)

: 
: 【 在 ihappy 的大作中提到: 】


ihappy (hungry christmas) 于Mon Dec 30 12:49:10 2002)
提到：

yep, NFL is not directly useful in specific ML algorithms. But i think it 
is important for knowing what is good ML algorithm and what makes good ML
algorithm.

【 在 daniel (飛翔鳥) 的大作中提到: 】
: 【 在 adson (自強乃報國之本) 的大作中提到: 】
: NFL just means you get some goodness but sacrifice others. It should 
: not be anticipated to get something good in every aspect or in every 
: way. This idea is not new. Even in 1980s', researchers have implicitly 
: regarded it as a norm. But until D. Wolpert, it has not been rigorously 
: justified. So, at present NFL theorem is regarded as the achievement 
: of him. It is worth noting that NFL does not conflict ML research because
: NFL just states that there is no universal winner, which is obvious
: correct. But for specific scenarios, there may exist a winner.
: for NFL, refer: 
: D.H. Wolpert and W.G. Macready. No free lunch theorems for optimization. 
: IEEE TEC97, 1(1)
: M. Koppen, D.H. Wolpert, and W.G. Macready. Remarks on a recent paper
: on the no free lunch theorems. IEEE TEC01, 5(3)


GzLi (笑梨) 于Mon Dec 30 16:38:02 2002)
提到：

I have a question.
daniel is excerllent at NN ensemble, can I do something on ML ensemble?
then NFL is not a truth.

【 在 ihappy (hungry christmas) 的大作中提到: 】
: yep, NFL is not directly useful in specific ML algorithms. But i think it 
: is important for knowing what is good ML algorithm and what makes good ML
: algorithm.
: 【 在 daniel (飛翔鳥) 的大作中提到: 】


daniel (飛翔鳥) 于Mon Dec 30 17:25:26 2002)
提到：

【 在 GzLi (笑梨) 的大作中提到: 】
: I have a question.
: daniel is excerllent at NN ensemble, can I do something on ML ensemble?
: then NFL is not a truth.

NN ensemble can be regarded as a branch of ensemble learning, 
but the techniques are almost the same. So, they are almost the
same thing, at least from some aspects. Note that Daniel mainly works
in machine learning

: 【 在 ihappy (hungry christmas) 的大作中提到: 】


GzLi (笑梨) 于Mon Dec 30 23:17:12 2002)
提到：

As I have known
you ensemble techniques are mainly on the same algorithm with 
defferent parameters, my question is that if there are any mixtures
of defferent algorithms i.e. boosting of bayesian network and SVMs.

and can you tell me what is the chinese term of boosting and ensemble ?
and bagging?
thank you!

【 在 daniel (飛翔鳥) 的大作中提到: 】
: 【 在 GzLi (笑梨) 的大作中提到: 】
: NN ensemble can be regarded as a branch of ensemble learning, 
: but the techniques are almost the same. So, they are almost the
: same thing, at least from some aspects. Note that Daniel mainly works
: in machine learning


adson (自強乃報國之本) 于Wed Jan  1 22:31:53 2003)
提到：

Thanks to daniel and ihappy.

Now I've gotten a general idea of NFL.


【 在 daniel 的大作中提到: 】

: 【 在 adson (自強乃報國之本) 的大作中提到: 】

: NFL just means you get some goodness but sacrifice others. It should 

: not be anticipated to get something good in every aspect or in every 

: way. This idea is not new. Even in 1980s', researchers have implicitly 

: regarded it as a norm. But until D. Wolpert, it has not been rigorously 

: justified. So, at present NFL theorem is regarded as the achievement 

: of him. It is worth noting that NFL does not conflict ML research because

: NFL just states that there is no universal winner, which is obvious

: correct. But for specific scenarios, there may exist a winner.

: for NFL, refer: 

: D.H. Wolpert and W.G. Macready. No free lunch theorems for optimization. 

: IEEE TEC97, 1(1)

: M. Koppen, D.H. Wolpert, and W.G. Macready. Remarks on a recent paper

: on the no free lunch theorems. IEEE TEC01, 5(3)



highso (漫步者) 于Sun Jan  5 12:55:53 2003)
提到：

the discussion has shift from the initial problem to the NFL
problem,:)
at my first post,I am puzzled at the application of the ML
to a practical application.From the discussion I can see that
the data preprocessing is important to the application.To a successful
solution of a specific problem,maybe domain knowledge contributed
to preprocessing and appropriate ML method should be given same attention

【 在 daniel (飛翔鳥) 的大作中提到: 】
: 【 在 adson (自強乃報國之本) 的大作中提到: 】
: NFL just means you get some goodness but sacrifice others. It should 
: not be anticipated to get something good in every aspect or in every 
: way. This idea is not new. Even in 1980s', researchers have implicitly 
: regarded it as a norm. But until D. Wolpert, it has not been rigorously 
: justified. So, at present NFL theorem is regarded as the achievement 
: of him. It is worth noting that NFL does not conflict ML research because
: NFL just states that there is no universal winner, which is obvious
: correct. But for specific scenarios, there may exist a winner.
: for NFL, refer: 
: D.H. Wolpert and W.G. Macready. No free lunch theorems for optimization. 
: IEEE TEC97, 1(1)
: M. Koppen, D.H. Wolpert, and W.G. Macready. Remarks on a recent paper
: on the no free lunch theorems. IEEE TEC01, 5(3)


GzLi (笑梨) 于Sun Jan  5 22:06:06 2003)
提到：

Yes, I think so.

【 在 highso (漫步者) 的大作中提到: 】
: the discussion has shift from the initial problem to the NFL
: problem,:)
: at my first post,I am puzzled at the application of the ML
: to a practical application.From the discussion I can see that
: the data preprocessing is important to the application.To a successful
: solution of a specific problem,maybe domain knowledge contributed
: to preprocessing and appropriate ML method should be given same attention
: 【 在 daniel (飛翔鳥) 的大作中提到: 】
: (以下引言省略 ... ...)