MPC-based ABR borrows from control theory: at each decision point, predict throughput for the next K segments, enumerate the K-segment rendition trajectories, score each by a utility function (bitrate − rebuffer penalty − switch penalty), pick the trajectory whose first rendition wins. K is usually 5–10 segments. The MPC paper (ACM SIGCOMM 2015) demonstrated that this lookahead approach outperformed both pure throughput-based and pure buffer-based ABRs on a wide range of network traces.
MPC's strength is that it can model both short-term throughput predictions and the cost of switching. It's also flexible — the utility function is configurable, so an operator can tune for "few switches at any cost" or "maximum average bitrate" or anything in between. The trade-off is computational cost: at each segment boundary, the player enumerates K-length trajectories across the rendition ladder. For 6 renditions and K=5 that's 7776 trajectories, which is computable but not free.
In production MPC ideas show up in Shaka Player's ABR (with a smaller lookahead than pure MPC), in Netflix's per-title ABR, and in Pensieve, which replaced MPC's predictive engine with a neural network. Pure academic MPC is less common in production; the practical pattern is "MPC structure with one-step lookahead and aggressive smoothing" rather than the full 10-segment search.
