Switching acosh_abs to acosh_nan as was done for the logs

This also removes the regularizing '1' that was present in the old
definition. Instead the test domain is shifted.

src/Core.jl

@@ -35,7 +35,7 @@ import .OperatorsModule:
     log10_nan,
     log1p_nan,
     sqrt_abs,
-    acosh_abs,
+    acosh_nan,
     neg,
     greater,
     greater,

src/Operators.jl

@@ -59,8 +59,9 @@ function log1p_nan(x::T)::T where {T<:Real}
     x <= T(-1) && return T(NaN)
     return log1p(x)
 end
-function acosh_abs(x::T)::T where {T<:Real}
-    return acosh(abs(x) + convert(T, 1))
+function acosh_nan(x::T)::T where {T<:Real}
+    x < T(1) && return T(NaN)
+    return acosh(x)
 end
 
 # Generics:
@@ -75,7 +76,7 @@ log_nan(x) = log(x)
 log2_nan(x) = log2(x)
 log10_nan(x) = log10(x)
 log1p_nan(x) = log1p(x)
-acosh_abs(x) = acosh(abs(x) + 1)
+acosh_nan(x) = acosh(x)
 
 function sqrt_abs(x::T)::T where {T}
     return sqrt(abs(x))

src/Options.jl

@@ -19,7 +19,7 @@ import ..OperatorsModule:
     log2_nan,
     log1p_nan,
     sqrt_abs,
-    acosh_abs,
+    acosh_nan,
     atanh_clip
 import ..EquationModule: Node, string_tree
 import ..OptionsStructModule: Options, ComplexityMapping
@@ -112,7 +112,7 @@ function unaopmap(op)
     elseif op == sqrt
         return sqrt_abs
     elseif op == acosh
-        return acosh_abs
+        return acosh_nan
     elseif op == atanh
         return atanh_clip
     end

src/SymbolicRegression.jl

@@ -39,7 +39,7 @@ export Population,
     log2_nan,
     log10_nan,
     log1p_nan,
-    acosh_abs,
+    acosh_nan,
     sqrt_abs,
     neg,
     greater,
@@ -130,7 +130,7 @@ import .CoreModule:
     log10_nan,
     log1p_nan,
     sqrt_abs,
-    acosh_abs,
+    acosh_nan,
     neg,
     greater,
     greater,

test/test_operators.jl

@@ -10,7 +10,7 @@ using SymbolicRegression:
     log2_nan,
     log10_nan,
     sqrt_abs,
-    acosh_abs,
+    acosh_nan,
     neg,
     greater,
     greater,
@@ -33,6 +33,8 @@ for T in types_to_test
     @test isnan(log2_nan(-val))
     @test abs(log10_nan(val) - log10(val)) < 1e-6
     @test isnan(log10_nan(-val))
+    @test abs(acosh_nan(val2) - acosh(val2)) < 1e-6
+    @test isnan(acosh_nan(-val2))
     @test neg(-val) == val
     @test sqrt_abs(val) == sqrt(val)
     @test mult(val, val2) == val * val2

test/test_tree_construction.jl

@@ -8,7 +8,7 @@ include("test_params.jl")
 x1 = 2.0
 
 # Initialize functions in Base....
-for unaop in [cos, exp, log_nan, log2_nan, log10_nan, relu, gamma, acosh_abs]
+for unaop in [cos, exp, log_nan, log2_nan, log10_nan, relu, gamma, acosh_nan]
     for binop in [sub]
         function make_options(; kw...)
             return Options(;
@@ -56,12 +56,16 @@ for unaop in [cos, exp, log_nan, log2_nan, log10_nan, relu, gamma, acosh_abs]
 
             Random.seed!(0)
             N = 100
-            if unaop in [log_nan, log2_nan, log10_nan]
+            if unaop in [log_nan, log2_nan, log10_nan, acosh_nan]
                 X = T.(rand(MersenneTwister(0), 5, N) / 3)
             else
                 X = T.(randn(MersenneTwister(0), 5, N) / 3)
             end
             X = X + sign.(X) * T(0.1)
+            if unaop == acosh_nan
+                X = X .+ T(1.0)
+            end
+
             y = T.(f_true.(X[1, :]))
             dataset = Dataset(X, y)
             test_y, complete = eval_tree_array(tree, X, make_options())