A closed-loop brain-machine interface system (BMI) was implemented using freely-moving rats. Instead of reproducing continuous natural limb movements as in many other BMI work, abstract supervisory control commands such as Go left, Go right, were extracted from neurons in the motor and premotor areas of the rats brain. The control output was thus formulated as a solution of a nonlinear support vector machine (SVM). Five male Sprague-Dawley rats were able to use such a BMI. Furthermore, evidence was found that the animal changed his behavior and neural activity during the use of the interface from the hand-control phase to the brain-control phase. The analysis showed that the animal adapted a subset of its neural activities to make his decisions more distinct in the SVM decision space from neural activities, which subsequently led to improved brain-control task performance. Two independent approaches, an SVM model sensitivity analysis and a model-free mutual information analysis, pointed to the same subset of neurons that were responsible for such changes.