Nondeterministic finite automata (NFA) are widely used in computer science fields, such as regular languages in formal language theory, high-speed network monitoring, image recognition, hardware modeling, or even in bioinformatic for the detection of the sequence of nucleotide acids in DNA. They are also used in regular mode checking, in string solving, in verification of pointer manipulating programs, for construction of linear arithmetic equations and inequalities, for decision in WS1S and WS2S logic, and many others. Automata minimization is a fundamental technique that helps to decrease resource claims (memory, time, or a number of hardware components) of implemented automata and speed up automata operations. Commonly used minimization techniques, such as state merging, transition pruning, and saturation, can leave potentially minimizable automaton subgraphs with duplicit language information. These fragments consist of a group of states, where the part of language of one state is piecewise covered by the other states in this group. The thesis describes a new minimization approach, which uses SAT solver, which provides information for efficient minimization of these so far nonminimizable automaton parts. Moreover, the newly investigated method, which only uses solver information and state merging, can minimize the automaton similarly and on automata with low transition count faster than a tool RABIT/Reduce, which uses state merging and transition pruning.