Scaling techniques for $epsilon$-subgradient methods