Function Approximation - Linear
For every (s,a) there is a feature vector r(s,a)
Let Q’(s,a) = fw(s,a) = < r(s,a) , w>
Gradient-descent:
Ñw fw(s,a) = r(s,a)
w := w + a [ rt+1+gQ(st+1, at+1) - Q(st, at) ] r(s,a)
Previous slide
Next slide
Back to first slide
View graphic version