In this master’s thesis we have dealt with randomized algorithms. Randomized algorithms are a type of algorithms that in at least one part of their execution make a random decision about the further course of execution. We first define two theoretical models. The algorithms from the first model randomly select a deterministic algorithm, which they then use at a given input, while the algorithms belonging to the second model make more random decisions during execution. An example of an algorithm built on the first model is the \(R_k\) protocol, and an example of an algorithm built on the second model is the randomized Quicksort. We then dealt with the classification of randomized algorithms with respect to the probability of error. Las Vegas algorithms are randomized algorithms that guarantee that every compute answer is correct. We have listed two subtypes for them: those that allow the answer "?" and those that do not allow it. We have proven that the subtypes of these algorithms can be transformed to one another, and we have analyzed the Random - select algorithm as an example of a Las Vegas algorithm that does not allow the "?" answer, and the LV\(_{10}\) protocol that allows it. One-sided error algorithms allow errors, but only if the probability of an error is less than \(\frac{1}{2}\) and only at the inputs that should be accepted. We have provided examples of one-sided error algorithms: Randomized Equality Protocol and Simplified Solovay - Strassen Algorithm. Bounded-error algorithms require that the error probability for each input is less than \(\frac{1}{2} - \epsilon\), where \(\epsilon\) is a fixed number from \( \left(0, \frac{1}{2}\right]\). We have analyzed the rate of error reduction of a bounded-error algorithm A with respect to the number of independent repetitions with the same input. We have come to the conclusion that the probability of error on bounded-error algorithms can be lowered arbitrarily, with a constant number of independent repetitions. Unbounded-error algorithms only require that the error probability is less than \(\frac{1}{2}\). An example of such an algorithm is the UMC protocol, which we have proven to satisfy the definition of an unbounded-error algorithm.