At Motorola, I developed a link adaptation algorithm for our rugged, industrial mesh Wi-Fi access points. The goal was to create a system that was both stable, providing consistent performance, and reactive, adjusting in real-time to changing conditions for devices in moving vehicles. The key to achieving this was to find the best data rate (the raw speed at which the data is transmitted through the radio channel).
My initial algorithm was effective but not optimal. Finding the best data rate proved to be a complex optimization challenge. The relationship was non-linear: while slower rates offered better resilience to noise, they also had longer transmission times, increasing their vulnerability to packet collisions. Counter-intuitively, this meant a much faster data rate could succeed where a medium one failed, simply because it occupied the channel for a shorter duration, reducing the probability of a collision. This complexity necessitated a full search of all possible rates to find the true optimum.
To address this, I designed a stochastic algorithm that allocated a probability to each data rate based on its predicted performance. This system reserved a small portion of the bandwidth (5%) to continuously explore different rates, ensuring the network was always self-optimizing.
To avoid latency issues with protocols like TCP/IP, rates were tested in small, structured batches. Each batch front-loaded the risk by testing exploratory rates first, but ensured a high probability of overall success by concluding with a 'safe' rate chosen from recent, reliable options.
In production networks, the average throughput tripled, with some links experiencing a 10-fold increase in performance. I achieved this while maintaining the same high levels of stability and reactivity as my previous algorithm.