function[score]=rmho_score(dag,nodes_sizes,train_data,validation_data,class_node_number)
%the Risk Minimization Hold Out score
    
    %fix dag and data to topological ordering constraint
    ordering=topological_sort(dag);
    nodes_sizes=nodes_sizes(ordering);
    class_node_number=find(ordering==class_node_number);
    train_data=train_data(ordering,:);
    validation_data=validation_data(ordering,:);
    dag=dag(ordering,ordering);
    
    %find number of nodes
    N=length(dag);
    
    %all nodes are discrete
    discrete_nodes=1:N;   
    
    %build a Bayesian network using the DAG structure
    bnet=mk_bnet(dag,nodes_sizes,'discrete',discrete_nodes); 
    
    %apply random CPDs for the network (these values do not effect results!)
    seed=7;
    rand('twister',seed);
    for c1=1:N
        bnet.CPD{c1}=tabular_CPD(bnet,c1);
        %bnet.CPD{c1}=tabular_CPD(bnet,c1,'prior_type','dirichlet','dirichlet_type','BDeu');   %alternative
    end
    
    %learn network parameters based on training data
    bnet=learn_params(bnet,train_data);
    
    %score the Bayesian network over the validation data (zero-one loss)
    %engine=jtree_inf_engine(bnet);               %alternative inference engine - junction tree inference engine
    engine=global_joint_inf_engine(bnet); 
    validation_size=size(validation_data,2);
    rate_counter=0;
    for c1=1:validation_size
        %evidence is the current validation sample, excluding the class variable
        evidence=validation_data(:,c1);
        evidence{class_node_number,1}=[];
        %enter evidence to the inference engine
        [engine2,loglik]=enter_evidence(engine,evidence);
        %calculate P(c|x1,..xn)
        marg=marginal_nodes(engine2,class_node_number);
        estimation=find(marg.T==max(marg.T));
        if estimation==validation_data{class_node_number,c1}
            rate_counter=rate_counter+1;
        end
    end
    score=1-rate_counter/validation_size;
end