1214 
MONITORING 
REFERENCES 
1. WiDROW, B., and Hoff, M. E. Adaptive Switch- 
ing Circuits. IRE Wescon Convention Record, Part 
4, pp. 96-104, 1960. 
2. WiDROW, B. Adaptive Sampled-Data Systems. Pro- 
ceedings, First International Congress IFAC, Mos- 
cow, pp. 406-411, 1960. 
3. Rosen, J. B. The Gradient Projection Method for 
Nonlinear Programming, Part I — Linear Con- 
straints. J. Soc. Indus. App. Math. 8, 1960. 
DISCUSSION 
P. G. Katona, Case Western Reserve Uni- 
versity, Cleveland : I would be interested to find 
out why you decided on the adaptive control 
model rather than pattern recognition tech- 
niques? And especially, it seems to me that 
correlation analysis would be useful. These can 
be usually characterized by expotential type of 
functions. I assume that you would just do 
what these analyses indicate. You might be 
able to get away with a much simpler technique. 
Dr. Thompson: You are right. In fact, we 
could probably do this by adapting only two 
parameters — delay and gain. Now, in fact, we 
approached the problem that way initially and 
then I built a small special purpose computer to 
do it. There's one hooker to that. If we lived 
in a world without noise that would work fine, 
but unfortunately we live with noise. Blood 
pressures fluctuate. They don't appear as per- 
fectly orderly functions. Therefore, we have to 
have some way to detect when a response is 
over. It's a simple response and, yes, it's an 
expotential, but you have to know when it's 
over to make a measurement to decide if you 
got the response you wanted. You can use cor- 
relation techniques and pattern recognition 
techniques but then you're getting into the need 
for a computer again. I mean, now we're talking 
about memory and we're talking about a lot of 
complications. We had the technique described 
here already available to us, and it gets around 
having to do the more complicated pattern 
recognition problems. 
Our Phase I control is simply a delay and a 
first order expotential model plus a way to find 
where you are in a field of noise. Now, there's 
more to it though. If you try to do Phase III 
control, that is where you try to move states 
around optimally, you are dealing with a more 
complicated problem, and we could ask how far 
one need go in the direction of optimum control. 
At present, I don't know. As a researcher, I 
think it's exciting, but as a practical physician, 
I'm not sure whether we need optimum control. 
But if you are going to try to approach optimum 
control with the idea to get the pertinent states 
to follow your input commands as closely as 
possible, then you probably have to know a little 
bit more about the system's dynamics than can 
be represented by only two data points. I didn't 
say why we actually used 21 data points. They 
happen to be five seconds apart and it takes 21 
X 5 on the average for this animal, the dog, to 
get through his response in the worse case. So 
we're trying to cover the worst case so we can 
model it. I think I've answered all your ques- 
tions. If not, I'd be happy to take another whack 
at it. 
W. J. Sacco, Edgewood Arsenal, Maryland: 
I would like to try to translate what you're say- 
ing into the language that I know for the general 
problem that you solved. Now, did you have 
anything else besides the delay that you were 
trying to identify as you were going along? 
Were you trying to identify other parameters 
besides delay? 
Dr. Thompson : And gain. 
Dr. Sacco: And gain, right. Okay, now, the 
delay — you considered that to be a random 
variable constantly changing? 
Dr. Thompson : Delay is capable of constant 
change. In fact, it changes very slowly over a 
long period of time. 
Dr. Sacco : So, you were doing systems identi- 
fication along with the control problem through- 
out? 
Dr. Thompson : Yes, our model is in fact a 
systems identifier. 
Dr. Sacco: Constantly? 
Dr. Thompson : Constantly ! I didn't go into 
the mathematics. I think it was my fourth slide 
(see equation 1), where you have all these 
weights and you're comparing the output of the 
model with the output of the plant. What you 
basically do is say, "I've got a weight vector. 
I'm going to increment that weight vector by 
delta weight vector," which says I'm going to 
