Minor improvements and bugfixes related to NelderMead (#1094)

Author: pkofod

docs/src/user/minimization.md

@@ -214,6 +214,12 @@ Defined for multivariate optimization:
 * `g_converged(res)`
 * `initial_state(res)`
 
+Defined for `NelderMead` with the option `trace_simplex=true`:
+
+* `centroid_trace(res)` (with `extended_trace=true`)
+* `simplex_trace(res)`
+* `simplex_values_trace(res)`
+
 ## Input types
 Most users will input `Vector`'s as their `initial_x`'s, and get an `Optim.minimizer(res)` out that is also a vector. For zeroth and first order methods, it is also possible to pass in matrices, or even higher dimensional arrays. The only restriction imposed by leaving the `Vector` case is, that it is no longer possible to use finite difference approximations or automatic differentiation. Second order methods (variants of Newton's method) do not support this more general input type.
 

src/api.jl

@@ -49,6 +49,7 @@ function simplex_trace(r::MultivariateOptimizationResults)
     [ state.metadata["simplex"] for state in tr ]
 end
 function simplex_value_trace(r::MultivariateOptimizationResults)
+    tr = trace(r)
     if !isa(r.method, NelderMead)
         throw(ArgumentError("There are no simplex values involved in optimization using $(r.method). Please use f_trace(...) to grab the objective values from the trace."))
     end

test/general/api.jl

@@ -32,6 +32,18 @@
     Optim.optimize(d2, initial_x, LBFGS())
 
     Optim.optimize(f, initial_x, NelderMead())
+    ne_res = Optim.optimize(f, initial_x, NelderMead(), Optim.Options(store_trace=true,))
+    @test_throws ErrorException Optim.simplex_trace(ne_res)
+    @test_throws ErrorException Optim.simplex_value_trace(ne_res)
+    @test_throws ErrorException Optim.centroid_trace(ne_res)
+    ne_res2 = Optim.optimize(f, initial_x, NelderMead(), Optim.Options(store_trace=true,trace_simplex=true))
+    Optim.simplex_trace(ne_res2)
+    Optim.simplex_value_trace(ne_res2)
+    @test_throws ErrorException Optim.centroid_trace(ne_res2)
+    ne_res3 = Optim.optimize(f, initial_x, NelderMead(), Optim.Options(store_trace=true,extended_trace=true,trace_simplex=true))
+    Optim.simplex_trace(ne_res3)
+    Optim.simplex_value_trace(ne_res3)
+    Optim.centroid_trace(ne_res3)
 
     Optim.optimize(d3, initial_x, Newton())