Learning hidden Markov models with geometrical constraints

摘要: Hidden Markov models (HMMS) and partially observable decision processes (POMDPS)form a useful tool for modeling dynamical systems. They are particularly representing environments such as road networks office buildings, which typical robot navigation planning. The work presented here is concerned with acquiring models. We demonstrate how domain-specific information consaaints can be incorporated into the statistical estimation process, greatly improving learned in terms of model quality, number iterations required convergence robustness to reduction amount available data. present new initialization heuristics used even when data suffers from cumulative rotational error, update rules parameters, an instance generalized EM, strategy enforcing complete geometrical consistency model. Experimental results effectiveness our approach both simulated real data, traditionally hard-to-learn environments.