In wireless communications,spatial modulation（SM）is a simple and promising multiple-input multiple-output（MIMO）technology to overcome the drawbacks existing in conventional MIMO systems for the inherent characteristics,e.g.,the activation of single transmit antenna per each symbol period to avoid inter-channel interference（ICI）and inter-antenna synchronization（IAS）.SM can also reduce the transmit power due to the activation of one transmit antenna and achieve high bit rate by additional bits conveyed by the activated antenna index.On the other hand,SM is very sensitive to channel fading and interference.To deal with this issue,adaptive modulation order selection（MOS）is adopted in SM,where different modulation orders can be adjusted to the available channel conditions to enhance the system bit error rate（BER）performance.However,the classical optimal adaptive MOS in SM is usually performed through conducting an exhaustive search（ES）over all the received symbols,which results in high system complexity.Recently,machine learning（ML）technique has proved the efficient reduction of complexity in SM-MIMO systems.In this thesis,we employ supervised learning,which is a ML task of learning a function that maps an input to an output based on a labeled training data.Specifically,we utilize supervised ML technique to select the optimal modulation order in SM and provide an effective complexity-performance tradeoff under the given spectral efficiency（SE）.The main contributions of this thesis focus on the following.Firstly,two supervised machine learning classifiers（MLCs）,K-nearest neighbors（KNN）and support vector machine（SVM）are proposed to convert the classical adaptive MOS optimization problem into a multi-class classification problem and predict the modulation order candidate which provides the maximum minimum Euclidean distance（MED）by training the relevant data.As a result,the system complexity is significantly reduced at the cost of marginal BER performance loss.Secondly,to further improve the tradeoff between the system complexity and the BER performance,this thesis proposes the supervised learning feed-forward artificial neural network（ANN）based adaptive MOS scheme.In this scheme,the ANN iteratively employs the learning parameters in an innovative way and allows the adaptive operations in training to achieve a high learning ability.Simulation results and complexity analysis validate that,for a given SE,all the proposed adaptive MOS schemes have much lower system complexity,where both KNN and SVM based adaptive MOS schemes achieve the sub-optimal BER performance,while ANN based adaptive MOS scheme attains the same BER performance as the optimal ES based adaptive MOS scheme.