This preview shows page 1. Sign up to view the full content.
Unformatted text preview: z ∂wi Substituting (3) and (4) into (2):
(5) ∂P ∂P ∂y i
= ( y * − y ) y (1 − y ) xi
∂wi ∂yi ∂wi Substituting (5) into the weight equation (1): wi' = wi + r( y * − y ) y (1 − y ) xi ... which can be written in terms of δ, which is the derivative of P with respect to total input, ∂P ∂y
= ( y * − y ) y (1 − y ) = y (1 − y )( y * − y )
into this equation.)
(Note that to change P, one only needs to substitute the new
wi' = wi + rδixi , where δi = ∂P
δi yi For an inner node, we use a similar equation that takes into account the output yi for that node, and the δ and
w values for the next layer, where "next layer" means the neighboring nodes one layer closer to the output
δi y = x δj y
y h = xi
wi' = wi + rδixi , where δi = yi(1 − yi) δjwij ∑ and wij is the weight between layers i and j. ... Summary: For each layer, wi' = wi + rδixi , where for outer layer δi = y (1 − y )( y * − y ) ∑δ w inner layer δi = yi(1 − yi) j ij 11...
View Full Document
This document was uploaded on 03/17/2014 for the course EECS 6.034 at MIT.
- Fall '10
- Artificial Intelligence