Implementation of AdaBoost algorithm

I know this is not a programming forum but I'm trying to implement the AdaBoost algorithm with linear SVMs as weak learners. I modified the SVM such that it can take weights into account. I also checked that the implementation of the weighted SVM works 100% correct.

So now I wanted to implement AdaBoost. This is the (Matlab) code:

    function models = adaboost( X, Y )
        N = length(X);
        G = 1/N * ones(N,1); %The weights
        M = 20; % Number of boosting iterations 
        models = cell( M, 4 ); % will keep track of learned parameters
        for m=1:M
            Cs = [.01,.1,.5,1,10]; %The cost parameters of the linear SVM
            best_error = -1;
            C_opt = -1;
            theta_opt = -1;
            b_opt = -1;
            %Learn the best parameter
            for cc=1:length(Cs)
                C = Cs(cc);
                [theta,b,S] = WSVM.wsvm( X, Y, C, G ); %So this here is a weighted SVM
                YPred = WSVM.classify( X, theta, b );
                %err = 1/sum(G) * sum( G.*(Y ~= YPred) ); %alternative as proposed in Bishop - Pattern recognition page 658
                err = sum(.5*G.*abs( Y - YPred ))/sum(G); %another way of error calculation but essentially equal to line above
                if err < best_error | best_error < 0
                    C_opt = C;
                    theta_opt = theta;
                    b_opt = b;
                    best_error = err;
                end
            end

            %Calculate the error again for the best parameter
            YPred = WSVM.classify( X, theta_opt, b_opt );
            err = sum(.5*G.*abs( Y - YPred ))/sum(G);
            alpha = log( (1-err)/err );

            %Update weights for the next round
            %G = G.*exp( alpha .* (YPred ~= Y) );
            G = G.*exp( - alpha.*Y.*YPred );
            G = G/norm(G);

            %Store model parameters
            models{ m, 1 } = theta_opt;
            models{ m, 2 } = b_opt;
            models{ m, 3 } = C_opt;
            models{ m, 4 } = alpha;
        end
end

I cannot see any error in my code and I think I the implementation corresponds correctly to the AdaBoost algorithm. But it does not work. This is a sample plot I get (first six leant SVMs)

See a plot here: AdaBoost - Plot of first 6 weak classifiers

The cyan colored line denotes the theta vector that describes the orientation of the separation plane (black), as one expects it is perpendicular.

The plot doesn't look good in my opinion...

EDIT:

Finally I come to the conclusion that my algorithm is not wrong at all. I've tested the dataset against WEKA using libSVM and the results are more or less equally poor. Then I tested the same dataset with DecisionStump and I hat an AUC of 94% versus 54% using libSVM or my implementation. Although I think I used a rather simple 2D toy dataset (see the plots of my adaboost implementation) I'm quite suprised that adaboost does not handle it appropriatly. Therefore I tested the algorithm on the banana dataset and it did not work for libSVM nor DecisionStump.

THEREFORE:

I come to the conclusion that AdaBoost is not a very good algorithm.