Figure 6

Cumulative throughput of the users for Markov channels, with a loss probability 0.05 at the “good” state and a loss probability of 0.5 at the “ba” state. The state transition probability is 0.01. In the greedy algorithm, the throughput increases at a constant rate, determined by the average loss probability seen by the user. However, the semi-greedy algorithm exploits the multi-user diversity, serving users with good channels while neglecting ones with bad channels. As a result, users with good channels experience a very steep increase in the reward, and the total system throughput is higher in the semi-greedy approach.