Learning of motor control using simulated brain base on biological signals

(1)我们认为，基于血红蛋白密度估计EMG和力振幅的可能性。在实验中，受试者进行了三个力幅水平的等长运动，以测量肌电，力幅和血红蛋白密度，并探讨这些关系。我们证实了这些测量之间的强相关性。从这些关系，我们提出了两个估计模型，一个是估计的肌电从血红蛋白密度和其他的是估计的力的振幅从估计的肌电。最后，我们展示了一个BMI系统应用估计模型来控制手臂机器人的例子。(2)本研究提出一种方法来分类近红外光谱（NIRS）信号响应左手或右手运动执行。在该方法中，我们选择的特征作为采样间隔和测量通道的近红外光谱信号。支持向量机被用作分类器。结果表明，该方法具有很高的泛化能力。结果表明，该方法可以有效地解决近红外光谱信号的两类分类问题。(3)我们将事件相关电位（ERP）应用于相当于奖励和惩罚信号的强化信号。我们进行了一次脑电图测试，志愿者们在测试中确定了一项任务的成功或失败。我们证实，有不同的脑电图取决于任务是否成功，并建议ERP可能被用作奖励的强化学习。我们使用SVM识别P300。我们选择了SVM中的特征向量，该特征向量由六个通道中每个通道的每50 ms的平均值组成，总共700 ms。我们可以建议使用P300的强化学习可以准确地执行。

(1) We consider a possibility for estimating EMG and force amplitude based on hemoglobin density. In experiments, subjects carried out isometric movements of three levels of force amplitude in order to measure EMG, force amplitude and hemoglobin density, and these relationships were investigated. We confirmed strong correlations between these measurements. From these relationships we propose two estimation models; one is to estimate the EMG from hemoglobin density and the other is to estimate the force amplitude from the estimated EMG. Finally, we show an example of a BMI system applying estimation models to control an arm robot.(2) This research proposes a method to classify near-infrared spectroscopy (NIRS) signals responded by left-hand or right-hand motor execution. In the proposed method, we selected features as sampling intervals and measured channels from NIRS signals. Support vector machines were used as a classifier. It is shown that the proposed method presents the high generalization capability. The result suggests that the proposed method can be useful to classify the two-category classification problem of NIRS signals.(3) We applied event-related potential (ERP) to reinforcement signals that are equivalent to reward and punishment signals. We conducted an EEG in which volunteers identified the success or failure of a task. We confirmed that there were differences in the EEG depending on whether the task was successful or not and suggested that ERP might be used as a reward of reinforcement leaning. We used a SVM for recognizing the P300. We selected the feature vector in SVM that was composed of averages of each 50 ms for each of the six channels for a total of 700 ms. We can suggest that reinforcement learning using P300 can be performed accurately.