Predicting packet loss statistics with hidden Markov models for FEC control

Media streaming applications often try to cope with packet losses using end-to-end recovery mechanisms like FEC. Predicting future losses is critical to choose the proper amount of redundancy needed to recover data. We propose a hierarchical model where the short-term dynamics of losses is driven by 2-state Markov chains while longer-term network changes (e.g., congestion) are modeled by a HMM. Based on this model, we develop two adaptive algorithms that predict future loss statistics and dynamically adjust FEC parameters. First, we predict loss rates and use these estimates to tune redundancy in Reed–Solomon codes. Second, we predict both loss rate and burstiness to select the optimal scheme among a set of parity-based FEC schemes. We perform experiments with packet loss traces to evaluate these algorithms, and compare their performance to standard approaches of FEC selection. Our results show that HMM-based prediction is more effective than other approaches, achieving higher quality improvements with small transmission overhead.