Learning internal representations by error propagation