Kalman filter model of adaptation Let us simulate an experiment in which we impose a perturbation on a movement, we reverse the perturbation for a few trials, and then impose an ‘error-clamp’. The target is at 90 deg, we generate a motor command u, and the experimenter imposes a 30 deg perturbation on our movement. After a few hundred trials, the rotation is reversed to -30 deg, and then an error-clamp is imposed on our movements. We assume that the learner makes the following assumptions: 1. There is a perturbation that is affecting my movements. That perturbation is driven by something that has multiple timescales. The perturbation is a sum of these hidden states. 2. The hidden states change from trial to trial. Some change slowly, others change fast. 3. The states that change fast have larger noise than states that change slow. u r1 h r2 ( n 1) x (n) Ax y (n) c x T (n) bu ε x (n) εy (n) εx N 0 ,Q εy N 0 , 2 y The perturbations that are produced by the experimenter: Initial conditions r1 x r2 h 0 b 0 1 0.995 A 0 0 0.01 Q 0 0 0 0.8 0 0 0 0 0 1 0 9 2 c 1 1 1 T x 0 0 0 P (0 ) 1 0 0 0 t 100 0 0 r 0 200 0 0 0 0 (n ) 0 30 30 0 , y yˆ n 1, , 300 n 301, , 500 n 501, , 520 n 600 (erro r clam p ) Using the Kalman filter approach, simulate the learner’s performance using the initial conditions given in the last slide. Plot u as a function of trial number. Plot y as a function of trial number. Plot the first and second element of the x vector (fast and slow states). Also plot the sum of these two states. Note the spontaneous recovery following start of error-clamp trials. t = 90 target is at 90deg u = t - d xˆ (n) T yˆ(n) = cT xˆ d = éê 1 1 0 ùú ë û T (n n-1) (n n-1) y (n) = u(n) + r (n) + e y xˆ (n n) = xˆ k = (n n-1) P (n) P xˆ (n n) cT P (n n-1) ( c c + s y2 ) = I - ck(n)T P (n+1 n) = Axˆ (n+1 n) = AP P ( +k(n) y (n) - yˆ (n) (n n-1) ( e y ∼ N 0,s y2 (n n-1) (n n) +bu(n) (n n) AT +Q ) ) 95 u x1 x2 (n) 20 x1 90 10 85 x2 80 0 75 10 70 65 0 100 200 300 400 500 600 700 200 300 400 500 600 700 120 (n) yˆ 100 80 60 40 0 100 0 100 200 300 400 500 600 700