Bill's Gate IIIc

From Mervyn van Kuyen (971001 23.52)

[Bill Powers]

What is the output of the neural net, and what does it do to its
surroundings? I'm just asking a general question, not asking about details
of design. As part of a behaving system, what role would this neural net play?

I think I already answer this question in the same post:

... in Fig 1, REF has to be generated by an independent, predefined
mechanism. Otherwise the system would 'die out' if it faced maximal input
(resulting in no error, no triggering of the network). So this network
can be used to produce _actions_ to minimize the detected mismatch.

This network produces action to minimize the detected error.

In Fig.2 the opposite is true: it detects ALL mismatch, like you've
said before as well, albeit WITHOUT sign. This problem can be solved
however by having the network generate the _reference signals_
as well as _actions_ to try to minimize the mismatch. This allows
it to know the sign perfectly well and exploit this knowledge,
as I wrote two days ago (in 'RE: Bill's Gate IIb'):

This network produces both actions and references for the same
purpose. Both have, in terms of their actions, the same goal
as PCT networks.

Do you see (considering my posts) how the 'absence of the sign'
problem can be overcome if the network has access (since it
produces them) to the references, where in PCT the reference
is effectively hidden for the network?

Regards,

Mervyn

[From Bill Powers (971002.0733 MDT)]

Mervyn van Kuyen (971001 23.52)--

Bill:

What is the output of the neural net, and what does it do to its
surroundings? I'm just asking a general question, not asking about details
of design. As part of a behaving system, what role would this neural net

play?

Mervyn:

I think I already answer this question in the same post:

... in Fig 1, REF has to be generated by an independent, predefined
mechanism. Otherwise the system would 'die out' if it faced maximal input
(resulting in no error, no triggering of the network). So this network
can be used to produce _actions_ to minimize the detected mismatch.

This network produces action to minimize the detected error.

In Fig.2 the opposite is true: it detects ALL mismatch, like you've
said before as well, albeit WITHOUT sign. This problem can be solved
however by having the network generate the _reference signals_
as well as _actions_ to try to minimize the mismatch. This allows
it to know the sign perfectly well and exploit this knowledge,
as I wrote two days ago (in 'RE: Bill's Gate IIb'):

This network produces both actions and references for the same
purpose. Both have, in terms of their actions, the same goal
as PCT networks.

If your system organizes itself to match the reference to the input at all
times, it doesn't ever need to act on its environment, does it?

You still haven't made it clear whether the comparison process you're
defining, or the action of the network, is analog (continuous) or digital
(on-off). When you speak of "maximal" input leading to zero action , I
deduce that you must be thinking in binary variables, because in an analog
control device "maximal" input would cause the input to exceed the
reference signal, and would lead to actions that tend to _reduce_ the input.

Do you see (considering my posts) how the 'absence of the sign'
problem can be overcome if the network has access (since it
produces them) to the references, where in PCT the reference
is effectively hidden for the network?

Well, the reference isn't exactly hidden from _all_ the networks inside the
organism, because the reference signal for one system is generated by the
output from another.

The trouble with "absence of sign" is that when there's a mismatch, the
output of your comparison elements doesn't indicate whether it's because
the reference is greater than the input, or the input is greater than the
reference -- so the network gets no information about which way to change
its output action. This means it must try one direction of change of
action, and if that doesn't work, try the other direction. So half of the
time the first move is in the wrong direction.

I'm sure you will say it doesn't have that problem, so probably the best
way to communicate what your system does is to show some actual application
of it to a control problem, like steering a car. Given actuators that can
turn the steering wheel and sensors that can detect the position of the car
relative to the road, how would your system work?

Best,

Bill P.