The paper is devoted to neural modeling of link occupancy distribution. Multi-service (i.e., bandwidth sharing between different traffic classes) models of a single, possibly wireless transmission link for rigid, adaptive and elastic traffic are developed based on Markov reward models. Link occupancy distribution is introduced as embedded, discrete time Markov chain researched with vector quantization. Link occupancy performance is simulated as a combination of single queues with random distributions of arrival processes and holding times in service phases. Link occupancy probability density is determined using learning vector quantization in a two-layered neural structure. Simulation and numerical results are shown.