 This paper proposes a novel approach to improving the performance of reinforcement learning-based brain-machine interface, RLBMI, systems. It uses a combination of temporal difference, TD, methods and attention-gated kernel reinforcement learning, CAGROL, to assign credits to different steps of the task, as well as discriminating between spatial locations in the reproducing kernel Hilbert space, RKHS. This allows for improved performance in complex BMI tasks, such as those involving multiple steps and providing feedback to the subject. This article was authored by Xiang Chen, Xiang Zhang, Yifan Huang and others.