Mitigating Estimation Bias with Representation Learning in TD Error-Driven Regularization