To address the issue of joint adaptation of rate, MIMO (multiple input multiple output) mode and channel width in IEEE 802.11n wireless networks, a joint adaptation algorithm based on non-stationary multi-armed bandit learning approach is proposed, and a novel reward function is also presented. To reduce the convergence time of the algorithm mentioned above, the prediction algorithms of MCS (modulation and coding scheme), MIMO mode and channel width based on classification and regression trees are developd to effectively utilize the statistical data collected by the wireless network interface driver to predict the reward values of different combination of MCS, MIMO mode and channel width, and shrink the search space of the joint adaptation algorithm. The proposed algorithm is easy to implement, approximately optimal, and has low computation complexity. The real experiment results show that the UDP throughput is improved significantly by the proposed algorithm under the interference-free environment and the environment with different interference conditions.