In digital mobile communication systems, intersymbol interference is one of the main causes of degrading system performance. Decision feedback equalization (DFE) is the commonly used remedy for this problem. Since the channel is fast-varying, an adaptive algorithm possessing a fast convergence property is then required. The least mean square (LMS) algorithm is well known for its simplicity and robustness; however, its convergence is slow. As a consequence, the LMS algorithm is rarely considered in this application. In this paper, we consider an LMS-based DFE for the North American IS-136 system. We propose an extended multiple-training LMS algorithm accelerating the convergence process. The convergence properties of the multiple-training LMS algorithm are also analyzed. We prove that the multiple-training LMS algorithm can converge regardless of its initial value and derive closed-form expressions for the weight error vector power. We further take advantage of the IS-136 downlink slot format and divide a slot into two subslots. Bidirectional processing is then applied to each individual subslot. The proposed LMS-based DFE has a low computational complexity and is suitable for real-world implementation. Simulations with a 900- MHz carrier show that our algorithm can meet the 3% bit error rate requirement for mobile speeds up to 100 km/hr.