Radial Basis Function Neural Network - the Algorithm

Step 0: read file of feature vector and label (target) data {(x^(q), t^(q)): q = 1,...,Q}

Step 1: draw random weights {w_mj: m=1,...,M, j=1,...,J} between 0.0 and 0.5

            set step sizes η = 0.5,, set stopping criterion ε = 0.000001

            select M centers in feature space so Gaussians cover it

            for m = 1 to M do

                compute σ_m = 1/(2M)^1/N

Step 2: compute {y_m: m=1,...,M} as given above in Equation (1)      //One-time computation

Step 3: compute {z_j: j=1,...,J} from Equation (3)

Step 4: compute E from above from Equation (2) and put Enew = E

Step 5: for m = 1 to M do                   //Adjust all weights by steepest descent

                 for j = 1 to J do

                         w_mj = w_mj - η(2/MQ)Σ_q=1,Q (z_j^(q) - t_j^(q))y_m^(q)

                        {if one of every 10^th iteration (if adjusting biases)

                              b_j = b_j - α(2/Q)Σ_q=1,Q (z_j^(q) - t_j^(q)) }

Step 6: put E = Enew

            compute Enew from Equation (2) above with new weights

            if Enew < E then η = 1.2η

            else η = 0.9η

Step 7: if Enew > ε

                  compute {z_j: j=1,...,J} from Equation (3)

                 goto Step 5

            else stop