Beam-Limited k-Step Lookahead for Computationally Efficient HMM Decoding

Hidden Markov Models (HMMs) are widely used in many sequential decision-making problems due to their ability to model time-related dependencies. The standard decoding methods in these models, such as the Viterbi algorithm, are limited by their dependence on past observations only. Thus, this leads to unpredictability when future information is available. In this work, we propose a decoding strategy called Beam-Limited k-Step Lookahead that looks k-step ahead as k-step Discreate Control Synthesis to make use of future information. The proposed method achieves a balance between decoding accuracy and computational complexity by constraining the search space to the top M most promising paths. Experimental results on synthetic HMM data show that our new decoding strategy significantly improves decoding accuracy over classical Viterbi decoding. The findings highlight the potential of this new strategy to improve the performance of sequential decoding systems.