nlp_008_011 has a better solution than the recorded one

[xhub]

It looks like the nlp_008_011 problem has a better solution than the recorded one. This was found by KNITRO:

> value.([x,y,z])
3-element Array{Float64,1}:
  0.006189119075609131
  3.717464829723498   
 -3.719719919556328

julia-1.2> value(y) - (exp(-value(x)-2) + exp(-value(z)-2) - 2)
-9.800737998943987e-11

julia-1.2> value(x)^2 - value(y)^2 - value(z)^2
-27.65582273498036

julia-1.2> value(y) - value(x)/2 - value(z)                                                                                                                
7.4340901897420215                                                                                                                                         
                                                                                                                                                                                                                                                                                                                                                                                                                     
julia-1.2> value(x) + value(y)^2 + value(z)^3                                                                                                              
-37.641487400482944 

There is a slight infeasibility for the 1rst constraint, other it all looks good. I check the model against BARON, Couenne, Antigone, and they all find this solution.

[ccoffrin]

@xhub thanks for reporting this. My recommendation would be to modify the test case so that all solvers find the same optimal solution. These tests should be "mathematically" trivial and focus only on testing the JuMP/MOI APIs.

A PR would be welcome and appreciated.

[odow]

Ditto with 008_010.

I'll think about changing the test.