Merge pull request #42 from vvpisarev/catch_domain_errors_linesearch

Catch domain errors linesearch

src/linesearch.jl

@@ -63,18 +63,36 @@ function strong_backtracking!(
     nbracket_max = 200
     @logmsg LSLogLevel "==BRACKETING THE MINIMUM=="
     for nbracket in 1:nbracket_max
-        y, grad = fdf(x0, α, d)
+        local y, grad
+        while true
+            try
+                y, grad = fdf(x0, α, d)
+                break
+            catch err
+                if err isa DomainError
+                    α_new = (α + α_prev) / 2
+                    if α_new == α_prev || α_new == α
+                        @warn "==BRACKETING FAIL, LAST STEP VALUE RETURNED==" α
+                        @logmsg LSLogLevel "==LINEAR SEARCH INTERRUPTED=="
+                        return zero(α)
+                    end
+                    α = α_new
+                else
+                    rethrow(err)
+                end
+            end
+        end
         g = dot(grad, d)
         Δyp = (g + g0) * α / 2 # parabolic approximation
         if abs(Δyp) < ϵ
-            @logmsg LSLogLevel "" """
+            @logmsg LSLogLevel """
                 Δyp = $(Δyp) (*)
                 Δy = $(y-y0)
             """
             Δy = Δyp
         else
             Δy = y - y0
-            @logmsg LSLogLevel "" """
+            @logmsg LSLogLevel """
                 Δyp = $(Δyp)
                 Δy = $(Δy) (*)
             """

src/optimize.jl

@@ -191,6 +191,7 @@ function solver(
     maxcalls=nothing,
     constrain_step=nothing,
 )
+    x0 = argumentvec(M)
     grad_tol = (isnothing(gtol) || gtol < 0) ? zero(eltype(x0)) : gtol
     convcond = isnothing(convcond) ? stopbygradient(grad_tol) : convcond
     iter_limit = (isnothing(maxiter) || maxiter < 0) ? typemax(Int) : convert(Int, maxiter)

test/interface.jl

@@ -40,5 +40,12 @@ end
             )
             @test optimize!(rosenbrock!, opt, init_vec) isa NamedTuple
         end
+
+        @testset "$(typeof(descent).name) (default parameters)" for descent in descent_methods
+            opt = Downhill.solver(
+                descent,
+            )
+            @test optimize!(rosenbrock!, opt, init_vec) isa NamedTuple
+        end
     end
 end

test/linesearch.jl

@@ -70,3 +70,13 @@ end
         @test isapprox(α, α₀, rtol=0.05)
     end
 end
+
+@testset "Limited domain" begin
+    function fdf(x, α, d)
+        z = x[] + α * d[]
+        return z - sqrt(z), [1 - 0.5 / sqrt(z)]
+    end
+
+    α = Downhill.strong_backtracking!(fdf, [0.5], [-1.0])
+    @test isapprox(α, 0.25, rtol=0.05)
+end