The packet error rate (PER) of wireless BICM-OFDM systems is notoriously difficult to predict analytically. This remains true even if all subcarriers use a common modulation and coding scheme (MCS). Link adaptation, which here shall be understood as the process of adapting the MCS in order to maximize goodput, therefore remains a major challenge. Non-parametric learning is an elegant way to evade the lack of robust analytical models. Learning from multidimensional features is particularly interesting because one-dimensional features can characterize frequency-selective channels only roughly. However, most of the literature discusses methods that are not truly online. Either the computational costs become unbearable over time or the method saturates and effectively stops learning. The modified k nearest neighbors algorithm (k-NN) seems to be the only exception currently. However, k-NN has well-known weaknesses in learning from small sample sets. Two adaptive kernel regression (AKR) methods are therefore proposed instead. Simulation results are reported for a setup in which several practically relevant conditions that have been mostly ignored in previous studies using multidimensional features (imperfect channel knowledge, Doppler shift, feedback delay, collisions) are modeled.