micro@
by minimizing sum |y-yhat|^3 ??? does it show any improvement? hard to imagine what improvement....
eaihua
abel
没想到还有宣传这个冬冬的！
虽然值得鼓励，但是感觉总是怪怪的
似乎还是利用平方和最小值作为判定标准来计算参数的吧？
似乎还是应该叫最小二乘吧，同时，经典ols已经可以处理提到的这些类模型了，模型稍微在数学形式上稍微改变一下就可以了，没有必要发明一个新名词。SPSS、SAS等等都可以处理这个冬冬了。
模型拟合以后的残差结构性质是建模和分析的关键所在，而不是其他什么
个人说法，不足采信，如有冒犯，多多包含。
yihui
做软件的这位老大应该有点设计软件的一般思想，起码来说，这个软件没考虑到“通用”，太小了！
eaihua
SPSS、SAS等处理复杂多维非线性问题误差大，你可以试一试。
yihui
刚刚我又过去看了一下，发现micro@和我都理解错了，不是min(y-yhat)^3，楼主的“最小三乘法”本质仍然是最小二乘法！呵呵，只不过是非线性的而已，y=b0+b1*x^k+error，解法依然是OLS，所以我认为楼主的这种名称不太妥当啊。
micro@
[quote]引用第7楼谢益辉于2006-11-04 22:05发表的“”:
刚刚我又过去看了一下，发现micro@和我都理解错了，不是min(y-yhat)^3，楼主的“最小三乘法”本质仍然是最小二乘法！呵呵，只不过是非线性的而已，y=b0+b1*x^k+error，解法依然是OLS，所以我认为楼主的这种名称不太妥当啊。[/quote]
Haha~~ I didn't check that page before. This new name cheated me!
This is non-linear least square, not least cube.
Hard to believe such things could be called "new" methods.
It's surely older than my age.
micro@
[quote]引用第6楼eaihua于2006-11-04 20:44发表的“”:
SPSS、SAS等处理复杂多维非线性问题误差大，你可以试一试。[/quote]
Please post a dataset on which you believe your software performs best, and let's try which one can get the smallest sum of squared error, using the same functional form for the mean and the same number of parameters. Hard to believe there's still much room to improve here.
eaihua
还是你自己找数试吧
micro@
haha, this is interesting. The one that says his method is better than existing ones doesn't provide any evidence, and asks others to believe his method is indeed better.
eaihua
yihui
It seems that neither of your examples provides the evidence
micro@
[quote]引用第13楼谢益辉于2006-11-06 20:08发表的“”:
It seems that neither of your examples provides the evidence [/quote]
Right. All of his examples are unfair. The form of expectation is different, and the number of parameters are different.
Model assessment is not simply how the model fits to the data, but also how complex the model is. If only looking at residual sum of squares (basicall, this is what's used in all his measures), the data themselves are the best estimates. That is, I can use n parameters to describe n data points. What's the use of that?
The better fit claimed by the author is simply due to using a larger model, since multiple regression is nested in that model.
What's more, the estimation uncertainty of the exponents is very hard to be accounted for when using the usual F-statistic.
liuxingyu
设计软件的人一定是个天才，厉害厉害！把天赋接济我点咋样？
jasonberg
先试用看看，
