The Controller Area Network (CAN) is a communication bus for message transaction in real-time environments. A real-time system typically consists of several classes of messages and a scheduler is responsible to allocate network resources to fulfill timing constraints. Given sufficient bandwidth, the static scheduling algorithms can meet the bounded time delay. However, due to the availability of network bandwidth, not all messages can gain enough bandwidth to achieve the best control performance. Therefore, adaptive message scheduling policies that optimize the bandwidth utilization while supporting timeliness guarantees are of special interest. In this paper, we devise an online adaptive Message Scheduling Controller (MSC), which is designed to dynamically respond to the network dynamics. The MSC is realized by the Radial Basis Function (RBF) network with supervised parameter tuning. Two novel learning algorithms provide complementary effects (1) to prevent possible causes of non-uniform bandwidth allocation and (2) to reduce possibilities of transmission failures. Simulation results show that the proposed MSC in conjunction with parameter adaptation outperforms the existing Fixed- Priority scheduling and Earliest-Deadline First method, in terms of convergence speed, schedulability, and transmission failure rates.