On stochastic learning in predictive wireless ARQ

Traditional automatic repeat request (ARQ) protocols are channel unaware. That is, they react to channel errors by simply retransmitting erroneous packets and do not proactively decide whether or not to transmit a packet in a given slot based on past channel conditions. Clearly, ARQ protocols operating in this mode are not energy efficient. For example, continuously retransmitting erroneous packets when the wireless channel is in deep fade would cause significant wastage of transmission energy. In this paper, we present a stochastic learning automaton-based wireless channel state aware ARQ protocol. The learning automaton learns to predict and track the time-varying wireless channel conditions based on past observations. A Markov chain model for the channel state transitions is used. No a priori knowledge about the state transition probabilities is required by this predictor. Stochastic convergence of the learning algorithm is proved. The proposed ARQ protocol utilizes the predictions to compute transmission/retransmission policies in an online fashion. No pilot (training) symbols are used by the protocol for channel state prediction thereby avoiding any energy wastage due to the transmission of these symbols. Simulation results show that depending on the channel memory significant energy savings can be attained when compared with standard ARQ protocols. We also discuss the transmission energy versus delay trade-off.