performance improvements

src/common.jl

@@ -223,23 +223,21 @@ end
 
 function _predictrBCM(node::GPSplitNode, x::AbstractMatrix)
     μ = zeros(size(x,1))
-    t = zeros(size(x,1))
-    β = zeros(size(x,1))
 
     gp = leftGP(node)
     s = diag(kernelmatrix(gp.kernel, x, x)) .+ getnoise(gp)
 
+    C = deepcopy(1 ./ s)
+
     for (k,c) in enumerate(children(node))
         μ_, t_ = _predictPoE(c, x)
-        β_ = 0.5 * (log.(s) - log.(inv.(t_)))
-        t[:] += β_.*t_
-        μ[:] += β_ .* t_ .* μ_
-        β[:] += β_
+        s_ = 1 ./ t_
+        β_ = 0.5 * (log.(s) - log.(s_))
+        C += (β_ .* t_) - (β_ ./ s)
+        μ += μ_ .* (β_ .* t_)
     end
-    z = (1 .- β) ./ s
-    t += z
-    t[t .<= 0] .= 1e-8
-    return μ ./ t, t
+
+    return μ ./ C, C
 end
 
 function predict(node::GPSplitNode, x::AbstractMatrix)