Brain-machine interfaces (BMIs) aim to restore lost functions of paralyzed patients by interpreting motor intents into control signals. In a brain control (BC) scenario, the user drives the neuroprosthesis to accomplish some specific tasks using their brain activities without any natural arm movement. However, time-variant neural recordings make the BMI challenging to be applied to clinical usage. The existing decoding algorithm adopts a neural network structure, which is prone to trap in the local optimum and cannot provide an efficient online update for the neural-to-kinematic mapping, resulting in unstable and inaccurate prosthetic control in a brain control task.

In this thesis, I propose a cluster kernel reinforcement learning-based Kalman filter (CKRL-KF) to avoid the local optim...[ Read more ]

Brain-machine interfaces (BMIs) aim to restore lost functions of paralyzed patients by interpreting motor intents into control signals. In a brain control (BC) scenario, the user drives the neuroprosthesis to accomplish some specific tasks using their brain activities without any natural arm movement. However, time-variant neural recordings make the BMI challenging to be applied to clinical usage. The existing decoding algorithm adopts a neural network structure, which is prone to trap in the local optimum and cannot provide an efficient online update for the neural-to-kinematic mapping, resulting in unstable and inaccurate prosthetic control in a brain control task.

In this thesis, I propose a cluster kernel reinforcement learning-based Kalman filter (CKRL-KF) to avoid the local optimum problem. The neural patterns are projected into Reproducing Kernel Hilbert Space (RKHS), which builds a universal approximation to guarantee the global optimum. I test the proposed algorithm on both simulation and real data. The simulation setup is first designed to explore the pros and cons of the CKRL-KF under different neural encoding conditions compared with the existing algorithm. Then I test the algorithm with data collected on a BC two-lever discrimination task in four rats. The results demonstrate that CKRL-KF has superior performance with respect to many evaluation metrics, including the testing accuracy, stability, convergence speed, response time and distance, which indicates that the proposed algorithm is promising to be used in clinical BMI applications.