>_<.

v, w <- small random number
p <- number of traning pattern pairs
k <- 1
E <- 0

Step 2. Set learning rate(a) and Emax

Step 3. For each traning pattern pair (X, D)

do step 4 - 8 until k = p

Step 4. compute output

NET_z = Xk inner_product Vk
Z = f(NET_z)

f(NET_z) = 1 / (1 + exp(-NET_z))                                    ; unipolar sigmoid
                (1 - exp(-NET_z)) / (1 + exp(-NET_z))           ; bipolar sigmoid

NET_y = Z inner_product Wt
Y = f(NET_y)

f(NET_y) = 1 / (1 + exp(-NET_y))                                    ; unipolar sigmoid
                (1 - exp(-NET_y)) / (1 + exp(-NET_y))           ; bipolar sigmoid

Step 5. Compute output error

E <- 0.5 * (Dk - Yk)^2 + E

Step 6. Compute error signal

delta_y = (d-y) * y * (1 - y)                             ; unipolar sigmoid
              0.5 * (d-y) * (1-y^2)                         ; bipolar sigmoid

delta_z = z * (1-z) * sumation of delta_y*W      ; unipolar sigmoid
             0.5 * (1-z^2) * sumation of delta_y*W ; bipolar sigmoid

Setp 7. Update weights

Wk+1 = Wk + delta_Wk
        = Wk + a * delta_y * Zk

Vk+1 = Vk + delta_Vk
       = Vk + a * delta_z * Xk

Step 8. Increase counter and goto Step 3

k <- k + 1

 

Step 9. Test stop condition

If E < Emax, stop

else E <- 0, goto Step 3