At Motorola, we produced rugged, industrial mesh Wi-Fi access points. I wrote a link adaptation algorithm that selects the speed at which data is transmitted by the radio that achieved the following:
Stability — links must demonstrate consistent performance over time and not break sporadically
Reactivity — links must adjust in real-time to changing conditions, like when devices are mounted in moving vehicles
My first algorithm did not achieve optimality — it was possible to find better links, but it required a search across all possible link options at the same time (my algorithm, and those in the literature, could only adjust parameters up and down based on current conditions). The reason a full search was needed is that slow link parameters do better in the presence of noise, but they are more sensitive to packet collisions. The effects of collisions are nonlinear, meaning fast-link parameters may be better than the slow ones, even though the medium ones are worse than both.
I devised a stochastic algorithm that allocated a probability to each link parameter based on a predicted performance. The allocation was designed to reserve a fixed percentage of the bandwidth (say, 5%) to explore new parameters. Parameters were grouped in batches of 5 to ensure success before the fifth packet was sent in order to place an upper bound on latency (which was necessary to accommodate higher-layer protocols such as TCP/IP). In production networks, the average throughput tripled, while some links showed a 10-fold increase in performance. Stability and reactivity remained as good as with my previous algorithm.