Optimal combination policies for error exponent maximization in social learning