Supercharging Bluetooth Performance
with Machine Learning

maximize energy efficiency. So, when packets cannot be recovered, the receiver sends an indication back to the transmitter in the form of what’s called a NACK. This can be thought of as the opposite of an ACK or an acknowledgement that a packet has been successfully received, and the transmitter would then resend the entire packet and do so repeatedly until the receiver receives the message correctly.

In noisy environments, this can be extremely inefficient and result in many retransmissions that end up reducing overall system throughput, increasing power consumption and overall packet latency. In many cases, even multiple retransmissions can’t transfer packets across and users simply lose their music stream entirely. Figure 3 demonstrates this scenario wherein the number of packets making it through are so few that there is virtually no music playback possible in a traditional system.

Figure 3: Bluetooth audio with significant packet loss

Are blind retransmissions the only option? Can the transmitter do better? What additional information will it need to do this? Machine learning is an intrinsic part of these solutions. Multi-bit feedback technology provides additional information to the transmitter regarding the failing packets. This allows the transmitter to tailor the retransmissions to correct the failing packets more reliably. What additional information is needed and how it is used in the retransmissions is optimized using data-driven machine learning techniques.

Using machine learning, even the poor reception that is shown in Figure 3 can be corrected to deliver clean music audio. Figure 4 shows a real-world example of how the audio looks after ML-based corrections.

Figure 4: Bluetooth audio with ML-based corrections

Diving under the hood and looking at over-the-air real-world packet transmission and reception, Figure 5 shows how ML helps Bluetooth maintain wireless link performance even as the environment changes.

Figure 5: ML helps maintain wireless link performance
click to enlarge

To understand Figure 5, let’s first look at the graph on the left. This shows the cumulative number of packets successfully received on the y-axis and time on the