Reliability-Focused Scheduling with (m, k)-firm Deadlines over Wireless Channels - A Reinforcement-Learning Approach

In wireless radio applications, the quality of an underlying wireless channel is important, however, we know of a few applications that can tolerate some losses. As an example, real-time applications like streaming voice or video do permit packet loss and still retain a bearable service. With respect to quality of service requirement, we embrace one concise method to distinguish between the allowed and forbidden loss patterns. This method is known as the (m; k)- rm deadlines; at least m out of k consecutive packets have to be successfully delivered to their destination. We consider a point to multi- point network with a known population of wireless communication terminals and with one base station periodically polling all terminals. Given limited network resources, a recovery from data losses in such arrangement might be very challenging under high error rates and with large number of nodes. In this thesis, we consider policies that improve quality of multiple periodic streams by retransmission of failed packets. The base scheduler decides which streams to serve with respect to the primary goal of minimizing violation of the stream's deadline. We introduce an algorithm from Reinforcement Learning theory and compare its performance to a few baseline scheduling policies with static channels and channels with time-varying characteristics. We found that although the Learning scheme introduces good performance it doesn't outperform a baseline technique which is based on immediate slot allocation decisions with respect to packet error rate of each stream