N. P. THOMPSON, B. WIDROW AND C. SCHADE 
1215 
multiply the error times a f enagle factor, which 
we call the convergence factor, times the present 
input vector, divided by the correlation of this 
input vector with all of the previous input 
vectors. What you're saying is: "See if this 
input is very unique ; if it is, we're going to use 
a lot of it. If it's not, we're going to use a smidge 
of it." We're going to change our weights now 
by some percentage of the total change we need. 
We don't adapt in one step. We do not take our 
weights from some error to no error in one step. 
If you do, you get into instabilities. The fact is 
we used a convergence factor of 0.4 which takes 
us about three to four passes through the filter 
before we are down to where we'd like to be. 
This provides stability. In general, it gives us 
no big problem. I'd be happy to give you the 
equations if you'd like to see them. 
The big problem, as you may recognize is 
taking the inverse of this model (see equation 
4). In fact, you have a problem because the 
denominator of this particular equation happens 
to be weight number one, which may be zero, so 
that you're in trouble right off the bat. In fact, 
we go through our model and find out which is 
the first weight that is out of the noise. We use 
this weight as weight one. We remember how 
many weights we took off and that's our delay. 
Then we take all the dynamics and put them 
back on. You must also understand that we are 
dreamers. For blood pressure control this model 
is probably a cannon shooting a fly, we hope 
some day to go to more complicated things. We 
can visualize, for example, using a system like 
this where one can generate an Nth dimensional 
space and do multiparameter control. 
For example, maybe you don't just want to 
give this patient a volume expander and then an 
inotropic agent or vice versa. To get from point 
A to point B, you may have to follow a rather 
circuitous course without going into some for- 
bidden zones. You see, we're really trying to 
think ahead. So this model is not just to resolve 
this one problem, but we hope that it will give 
us the background to tackle more erudite prob- 
lems later on. 
Chairman: I think this is a very exciting 
paper and I have another question I'd like to 
ask you. In starting a new dog, or a new case, 
do you start with what you learned from the 
previous one or do you start from scratch again? 
It seems to me you could start with some sort of 
general average of where you were. 
Dr. Thompson : Yes, that was what I was — 
in my befuddled way — trying to point out. The 
model that we started out with was shown in 
the slide of the weights converging which had 
a peak, this is also the impulse response. That 
began with the model's weights for "our average 
dog." Now admittedly it was a poor one for this 
particular dog, but in general it's very close to 
where we're going, so that we don't have to 
change things very far, which cuts down our 
initial learning period. We could actually start 
anywhere. 
K. B. Larson, Washington University, St. 
Louis : Have you attempted to devise any biolog- 
ical models of the parasympathetic and sym- 
pathetic nervous systems and their different 
input-output response such that your adaptive 
process could shed light on? 
Dr. Thompson : No, we in fact avoided that 
question, because I think that's an entirely 
different question. There is one thing, along the 
medical line that we, of course, thought about 
and it was inherent in what I said just a moment 
ago. Maybe these models can give us some clues 
in pharmcodynamics, in the dynamics of drugs 
used in a multiparameter sense. That's biological 
modeling, if you wish. But trying to find out 
what the parasympathetic or sympathetic nerv- 
ous systems' roles are and so forth, no. We've 
made no attempt to do this. 
