Figure 1.

Model components and behavior. The user is composed of a generative model (1) that describes the probability of drawing a certain n-dimensional action q_k at the instant k. The action q_k contributes to generating a smooth trajectory of control signals s_k for the interface (2). The interface map computes a vector p_k of a dimension m < n and provides it to the user as feedback related to the smooth control action s_k (3). The user model processes the feedback assigning a reward for the action based on the feedback (4). After each iteration, the map recursively updates its parameters based on the distribution of the observed user commands s_k with learning rate γ. The user updates its generative model with learning rate η after a feedback is received, reinforcing the generated smooth action according to its reward.