The assurance of transmission quality in computer networks appears to be an important issue, particularly from the perspective of rapid development of computer networks. In the previous research there were many attempts to implement various Quality of Service (QoS) techniques. Unfortunately, QoS parameters are not always assured - most frequently user pays certain amount of money for transmission parameters which will never be achieved. This paper presents simulation of a new concept, which is determining the transmission quality with the application of Machine Learning (ML). In general, transmission quality is described by means of four parameters, i.e. bandwidth, delay, jitter and packet loss ratio. Pay&Require was suggested as a solution, which allows the assurance of transmission quality in computer networks. This purpose was achieved by the use of multi-agent system which monitors the transmission parameters and checks if they meet the customer’s expectations. The transmission quality rating is a significant factor of Pay&Require. ML was applied in the process of simulation and for the research purpose the assessment system of the transmission quality was implemented. It enabled the test users to assess the quality of transmission. Data obtained in such way was then used in ML classification. Simulations were performed for nine classifiers: Nu-Support Vector Classifier (Nu-SVC), k-Nearest Neighbors algorithm (kNN), Random Forest Classifier, C-Support Vector Classifier (C-SVC), Radius Neighbors Classifier, Nearest Centroid Classifier, Extra Trees Classifier, Decision Tree Classifier and Linear Support Vector Classifier (Linear SVC). Simulations were also performed for two variants of Stacking Classifier. The first variant was a combination of Linear SVC, C-SVC, Nearest Centroid and kNN as estimators and Logistic Regression as the final estimator. In the second variant Random Forest, Extra Trees and kNN were used as estimators and Logistic Regression was applied as the final estimator. The best classification result with respect to the tested data, was achieved by variant 1 Stacking Classifier, it had 89% sensitivity (overall accuracy), with 11/100 incorrect classifications.