Merge 9ad8b40 into 7523459

src/Population.jl

@@ -1,3 +1,4 @@
+using Random
 using FromFile
 @from "Core.jl" import Options, Dataset, RecordType, stringTree
 @from "EquationUtils.jl" import countNodes
@@ -41,30 +42,30 @@ Population(X::AbstractMatrix{T}, y::AbstractVector{T}, baseline::T;
 
 # Sample 10 random members of the population, and make a new one
 function samplePop(pop::Population, options::Options)::Population
-    idx = rand(1:pop.n, options.ns)
+    idx = randperm(pop.n)[1:options.ns]
     return Population(pop.members[idx])
 end
 
 # Sample the population, and get the best member from that sample
 function bestOfSample(pop::Population, options::Options)::PopMember
     sample = samplePop(pop, options)
-    if options.probPickFirst == 1.0
-        best_idx = argmin([sample.members[member].score for member=1:options.ns])
-        return sample.members[best_idx]
+
+    scores = [sample.members[member].score for member=1:options.ns]
+    p = options.probPickFirst
+
+    if p == 1.0
+        chosen_idx = argmin(scores)
     else
-        sort_idx = sortperm([sample.members[member].score for member=1:options.ns])
-        # Lowest comes first
-        k = range(0.0, stop=options.ns-1, step=1.0) |> collect
-        p = options.probPickFirst
+        sort_idx = sortperm(scores)
+        # scores[sort_idx] would put smallest first.
 
-        # Weighted choice:
+        k = collect(0:(options.ns - 1))
         prob_each = p * (1 - p) .^ k
         prob_each /= sum(prob_each)
         cumprob = cumsum(prob_each)
-        chosen_idx = findfirst(cumprob .> rand(Float32))
-
-        return sample.members[chosen_idx]
+        chosen_idx = sort_idx[findfirst(cumprob .> rand())]
     end
+    return sample.members[chosen_idx]
 end
 
 function finalizeScores(dataset::Dataset{T},

src/SymbolicRegression.jl

@@ -68,7 +68,7 @@ using Reexport
 @from "MutationFunctions.jl" import genRandomTree, genRandomTreeFixedSize
 @from "LossFunctions.jl" import EvalLoss, Loss, scoreFunc
 @from "PopMember.jl" import PopMember, copyPopMember
-@from "Population.jl" import Population, bestSubPop, record_population
+@from "Population.jl" import Population, bestSubPop, record_population, bestOfSample
 @from "HallOfFame.jl" import HallOfFame, calculateParetoFrontier, string_dominating_pareto_curve
 @from "SingleIteration.jl" import SRCycle, OptimizeAndSimplifyPopulation
 @from "InterfaceSymbolicUtils.jl" import node_to_symbolic, symbolic_to_node

test/test_prob_pick_first.jl

@@ -0,0 +1,35 @@
+using SymbolicRegression, Test
+
+n = 10
+
+options = Options(
+    binary_operators=(+, -, *, /),
+    unary_operators=(cos, sin),
+    probPickFirst=0.999,
+)
+
+for reverse in [false, true]
+    members = PopMember{Float32}[]
+
+    # Generate members with scores from 0 to 1:
+    for i=1:n
+        tree = Node("x1") * 3.2f0
+        score = Float32(i-1)/(n-1)
+        if reverse
+            score = 1 - score
+        end
+        push!(members, PopMember(tree, score))
+    end
+
+    pop = Population(members, n)
+
+    best_pop_member = [
+        SymbolicRegression.bestOfSample(pop, options).score
+        for j=1:100
+    ]
+
+    mean_value = sum(best_pop_member)/length(best_pop_member)
+
+    # Make sure average score is small
+    @test mean_value < 0.1
+end

test/unittest.jl

@@ -227,4 +227,9 @@ for fnc in [
 
     zero_tolerance = 1e-6
     @test all(abs.(test_y .- true_y)/N .< zero_tolerance)
-end
+end
+
+
+println("Testing whether probPickFirst works.")
+include("test_prob_pick_first.jl")
+println("Passed.")