0
$\begingroup$

Updated I Edited my code with the help of answer of Okkes Dulgerci as

ClearAll["Global`*"]
SeedRandom[];
ClearSystemCache[]
f1[a_, b_, g_] = 
  ProbabilityDistribution[
   3 a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 
     1) (1 - (1 + g x^(-b))^(-a/g))^2, {x, 0, \[Infinity]}];
f2[a_, b_, g_] = 
  ProbabilityDistribution[
   2 a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 
     1) (1 - (1 + g x^(-b))^(-a/g)), {x, 0, \[Infinity]}];
f3[a_, b_, g_] = 
  ProbabilityDistribution[
   a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 1) , {x, 0, \[Infinity]}];
t1 = RandomVariate[f1[3, 3, 2], {50, 25}];
t2 = RandomVariate[f2[3, 3, 4], {50, 25}];
t3 = RandomVariate[f3[3, 3, 6], {50, 25}];
lnL[g1_?NumberQ, g2_?NumberQ, g3_?NumberQ, a_?NumberQ, b_?NumberQ] := 
  Module[{n = 25, k = 3},
   n Log[Factorial[k]] + n k Log[a] + n k Log[b] - (b + 1) ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t1[[j, i]])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t2[[j, i]])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t3[[j, i]])\)])\)\)) - (\!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g1\)] + 1)\) Log[\((1 + g1\ 
\*SuperscriptBox[\((t1[[j, i]])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g2\)] + 1)\) Log[\((1 + g2\ 
\*SuperscriptBox[\((t2[[j, i]])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g3\)] + 1)\)\ Log[\((1 + g3\ \*
SuperscriptBox[
RowBox[{"(", 
RowBox[{"t3", "[[", 
RowBox[{"j", ",", "i"}], "]]"}], ")"}], 
StyleBox[
RowBox[{"-", "b"}],
FontWeight->"Plain"]])\)])\)\))
    + ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((\((k - 1)\)\ Log[\((1 - 
\*SuperscriptBox[\((1 + g1\ 
\*SuperscriptBox[\((t1[[j, i]])\), \(-b\)])\), 
FractionBox[\(-a\), \(g1\)]])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((\((k - 2)\) Log[\((1 - 
\*SuperscriptBox[\((1 + g2\ 
\*SuperscriptBox[\((t2[[j, i]])\), \(-b\)])\), 
FractionBox[\(-a\), \(g2\)]])\)])\)\))];
Table[FindMaximum[
  lnL[g1, g2, g3, a, 
   b], {{g1, 2}, {g2, 4}, {g3, 6}, {a, 3}, {b, 3}}], {j, 1, 50}]

Now this code works well. But

1) It gives some time initial values as an estimates (I think when not convergent). Can we block them?

2) Whole process repeat 50 times (using table command), is correct?.

Improve this question
$\endgroup$
10
  • $\begingroup$ What are data1, data2, data3? $\endgroup$
    – OkkesDulgerci
    Commented Mar 10, 2018 at 20:40
  • $\begingroup$ Data1, data2, and data3 are random varieties of custom distribution define above as f1,f2 and f3. $\endgroup$
    – SAAN
    Commented Mar 11, 2018 at 0:57
  • $\begingroup$ n = Length[t1]=5 why do you generate $5\times25$ data if don't use them all? $\endgroup$
    – OkkesDulgerci
    Commented Mar 11, 2018 at 4:30
  • $\begingroup$ Your distributions look awfully like BetaPrimeDistribution[]. $\endgroup$
    – J. M.'s missing motivation ♦
    Commented Mar 11, 2018 at 6:03
  • $\begingroup$ @OkkesDulgerci Actually I have to estimate the parameters 10000 time. Yes here I have to used n=25, 5 is incorrect. $\endgroup$
    – SAAN
    Commented Mar 11, 2018 at 14:43

1 Answer 1

Reset to default
1
$\begingroup$

I have tried to simplified your code. You don't need to maximize Log-likelihood, in Mathematica you can maximize likelihood. Please double check your likelihood. For this set up it does not work.

ClearAll["Global`*"]

f1[a_, b_, g_] =   
  ProbabilityDistribution[  
   3 a b x^(-b - 1) (1 + g x^(-b))^(-a/g -  
       1) (1 - (1 + g x^(-b))^(-a/g))^2, {x, 0, \[Infinity]}];

f2[a_, b_, g_] =   
  ProbabilityDistribution[ 
   2 a b x^(-b - 1) (1 + g x^(-b))^(-a/g -  
       1) (1 - (1 + g x^(-b))^(-a/g)), {x, 0, \[Infinity]}];

f3[a_, b_, g_] =   
  ProbabilityDistribution[
   a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 1) , {x, 0, \[Infinity]}];

data1 = RandomVariate[f1[1.5, 2, 0.2], {5, 25}];
data2 = RandomVariate[f2[1.5, 2, 0.4], {5, 25}];
data3 = RandomVariate[f3[1.5, 2, 0.6], {5, 25}];

lnL[g1_?NumberQ, g2_?NumberQ, g3_?NumberQ, a_?NumberQ, b_?NumberQ] := 
 Module[{n = Length[data1], k = 3},
     n Log[Factorial[k]] + n k Log[a] + n k Log[b] - (b + 1) ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((Log[\((data1[\([i]\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((Log[\((data2[\([i]\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((Log[\((data3[\([i]\)])\)])\)\)) - (\!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(-a\), \(g1\)] + 1)\) Log[\((1 + g1\ 
\*SuperscriptBox[\((data1[\([i]\)])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(-a\), \(g2\)] + 1)\) Log[\((1 + g2\ 
\*SuperscriptBox[\((data2[\([i]\)])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(-a\), \(g3\)] + 1)\)\ Log[\((1 + g3\ \*
SuperscriptBox[
RowBox[{"(", 
RowBox[{"data3", "[", 
RowBox[{"[", "i", "]"}], "]"}], ")"}], 
StyleBox[
RowBox[{"-", "b"}],
FontWeight->"Plain"]])\)])\)\))
       + ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
        1\), \(n\)]\((\((k - 1)\)\ Log[\((1 - 
\*SuperscriptBox[\((1 + g1\ 
\*SuperscriptBox[\((data1[\([i]\)])\), \(-b\)])\), 
FractionBox[\(-a\), \(g1\)]])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
        1\), \(n\)]\((\((k - 2)\) Log[\((1 - 
\*SuperscriptBox[\((1 + g1\ 
\*SuperscriptBox[\((data2[\([i]\)])\), \(-b\)])\), 
FractionBox[\(-a\), \(g2\)]])\)])\)\))]



NMaximize[ First@lnL[g1, g2, g3, a, b], {g1, g2, g3, a, b}, 
 Method -> "DifferentialEvolution", MaxIterations -> 10000]

Edit This works. I am not sure all parameters are positive!! Replace 2 by 10000 in Table

Table[f1[a_, b_, g_] = 
    ProbabilityDistribution[
      3 a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 
            1) (1 - (1 + g x^(-b))^(-a/g))^2, {x, 0, \[Infinity]}];
 f2[a_, b_, g_] = 
    ProbabilityDistribution[
      2 a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 
            1) (1 - (1 + g x^(-b))^(-a/g)), {x, 0, \[Infinity]}];
 f3[a_, b_, g_] = 
    ProbabilityDistribution[
      a b x^(-b - 1) (1 + g x^(-b))^(-a/g - 1) , {x, 0, \[Infinity]}];
 t1 = RandomVariate[f1[3, 3, 2], 25];
 t2 = RandomVariate[f2[3, 3, 4], 25];
 t3 = RandomVariate[f3[3, 3, 6], 25];
 lnL[g1_?NumberQ, g2_?NumberQ, g3_?NumberQ, a_?NumberQ, b_?NumberQ] := 
    Module[{n = 25, k = 3},
      n Log[Factorial[k]] + n k Log[a] + n k Log[b] - (b + 1) ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t1[\([i]\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t2[\([i]\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
          1\), \(n\)]\((Log[\((t3[\([i]\)])\)])\)\)) - (\!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g1\)] + 1)\) Log[\((1 + g1\ 
\*SuperscriptBox[\((t1[\([i]\)])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g2\)] + 1)\) Log[\((1 + g2\ 
\*SuperscriptBox[\((t2[\([i]\)])\), \(-b\)])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 1\), \(n\)]\((\((
\*FractionBox[\(a\), \(g3\)] + 1)\)\ Log[\((1 + g3\ \*
SuperscriptBox[
RowBox[{"(", 
RowBox[{"t3", "[", 
RowBox[{"[", "i", "]"}], "]"}], ")"}], 
StyleBox[
RowBox[{"-", "b"}],
FontWeight->"Plain"]])\)])\)\))
        + ( \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((\((k - 1)\)\ Log[\((1 - 
\*SuperscriptBox[\((1 + g1\ 
\*SuperscriptBox[\((t1[\([i]\)])\), \(-b\)])\), 
FractionBox[\(-a\), \(g1\)]])\)])\)\) + \!\(
\*UnderoverscriptBox[\(\[Sum]\), \(i = 
         1\), \(n\)]\((\((k - 2)\) Log[\((1 - 
\*SuperscriptBox[\((1 + g2\ 
\*SuperscriptBox[\((t2[\([i]\)])\), \(-b\)])\), 
FractionBox[\(-a\), \(g2\)]])\)])\)\))];
 NMaximize[ {lnL[g1, g2, g3, a, b], g1 > 0, g2 > 0, g3 > 0, a > 0, 
   b > 0}, {g1, g2, g3, a, b}], 2]

{{-85.6331, {g1 -> 2.64838, g2 -> 3.23484, g3 -> 8.12382, a -> 3.38133, b -> 2.34444}}, {-63.0394, {g1 -> 2.14105, g2 -> 4.01569, g3 -> 9.34059, a -> 3.05859, b -> 3.05695}}}

Improve this answer
$\endgroup$
2
  • $\begingroup$ ...well, the point of optimizing log-likelihood is that it's less likely to overflow or underflow while searching parameter space with it. $\endgroup$
    – J. M.'s missing motivation ♦
    Commented Mar 11, 2018 at 4:48
  • $\begingroup$ I know that. We are not sure if OP took log of likelihood correctly. $\endgroup$
    – OkkesDulgerci
    Commented Mar 11, 2018 at 4:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.