Init. |
Choose the kernel type and parameters.
Set the
M and
.
Set the set to empty set. Set (,C), (a,,P), and (K,Q) to empty values. t = 0 | |||||
Iterate | For a selected example (yt + 1,xt + 1) do | |||||
(a) |
Par. adj. (EP)
|
If
0 then (subtract contribution from
previous iteration)
| ||||
v_t+1^-1 | = | &lambda#lambda;_t+1^-1 - p ^T_t+1 K p _t+1 - p _t+1^T K C K p _t+1 | ||||
#<19103#> &alpha#alpha; | = | &alpha#alpha; + h _t+1 v_t+1 ( &alpha#alpha; ^T K p _t+1 - a_t+1) | ||||
#<19115#> C | = | C + v_t+1 h _t+1 h _t+1^T | ||||
(b) |
Online coeff.
|
Compute q(t + 1) and r(t + 1) using with (,), the first and second derivatives of:
| ||||
(c) |
Scalars & Vectors
|
| ||||
#<19156#> e _t+1 = Q k _t+1 | &gamma#gamma;_t+1 = k^*_t+1 - k _t+1^T#<19164#> e _t+1 | |||||
&sigma#sigma;^2_t+1 = k^*_t+1 + k _t+1^T C k _t+1 |
&eta#eta;_t+1 = (1 +&gamma#gamma;_t+1r^(t+1))^-1
| |||||
(d) |
EP update
|
| ||||
&lambda#lambda;_t+1 | = | - ( (r^(t+1))^-1 + &sigma#sigma;_t+1^2 )^-1 | ||||
(e) | Geom. test |
If
< then (perform full update)
| ||||
p _t+1 | = | #<19186#> e _t+1
otherwise(perform sparse update)
| ||||
Q ^new | = | Q + &gamma#gamma;^-1_t+1
(#<19209#> e _t+1 - e _t+1)(#<19213#> e _t+1+ e _t+1)^T
compute
| ||||
p _t+1 | = | e _t+1 (where pi is the i-th row of matrix P.) | ||||
(f) | par. update |
| ||||
C ^new | = | #<19245#> C + r^(t+1)&eta#eta;_t+1 s _t+1 s _t+1^T | ||||
(g) | removal | If (remove a ) | ||||
C ^new | = | C ^(r) + Q ^* Q ^*Tq^* - ( Q ^*+ C ^*)( Q ^*+ C ^*)^T q^* + c^* | ||||
Q ^new | = | Q ^(r) - Q ^* Q ^*Tq^* | ||||
P ^new | = | P ^(r) - P ^* Q ^*Tq^* where P(r) is the matrix P without the j-th column (or the last one, if we reordered the set), and P* is the column vector containing the projection coefficients corresponding to the j-th element. | ||||
(h) | Goto (a) |
t = t + 1
Select a new input (yt + 1,xt + 1) to process. Restart iteration. |
1