[Martin Taylor 2017.11.28.17.30]
Bruce, I believe we are confounding two issues. Rick originally claimed that "PCT" (presumably the true-blue authenticated by Rick version) does not have an "avoidance" control operation and separately that it does not have a choice function. I think those two issues should be addressed independently.
[From Bruce Nevin (2017.11.28 17:18 ET)]
...
You rejected Martin's list of examples, saying that none of them was a conflict situation. I believe that implicitly they are. Avoiding bumping into someone only occurs when control of some other variable(s) puts you in a place or on a path where collisions are likely (unless you control avoiding them).
Not necessarily at all. Let's look at the list again:
Imagine the following situations, all of which might plausibly complete a sentence that starts with “I want to avoid ��?
� bumping into anyone in a crowd
You could say that there is a conflict between getting where you want to go and wanting to bump into someone in the crowd, but why would you want to bump into someone and at the same time want to avoid bumping into them? Is it a conflict when you want to go through a door while avoiding bumping into the door jamb?
� falling into the old mineshaft in the field.
What is the conflict here? Do you imagine someone wanting the thrill of falling while at the same time wanting not to fall?
� falling over the balustrade on the seventh flooor balcony.
Same question.
� seeing the wine glass too near the edge of the table.
Here, and in a lot of them, what seems to be consistent is an imagination of some unwanted consequence that might (or might not) happen if you fail to avoid the reference value listed. In this case one might be imagining the wine glasses being inadvertently nudged and falling off.
� hearing foreigners talking their disgusting lannguage in the bus.
You might arrange that by ensuring that your bus route is chosen to avoid places where foreigners are likely to be, or by wearing effective earplugs, but where's the conflict?
� offending that person with whose policies I dissagree.
I can see a possible conflict here, because you may well also want to offend them.
� having that wall red when we redecorate.
What conflict? And so on for the rest of them. What is the to-be-approached perception and reference value that sets up the conflict?
� being in the same room with Jack.
� having Rachael see me with Dora.
� making a foot-fault when I serve in tennis.
� having Rachael be within talking distance of Doora.
� making a burning smell when I cook.
� being served a food to which I am allergic
� seeing the present government re-elected.
� being near someone smoking.
All of these have one thing in common, that there is no specific preferred alternative to the environmental condition that is to be avoided -- nothing to be approached. In none of them (possibly except one) is there any obvious conflict. The question is how do you model avoidance. Jeff has a suggestion which I have also proposed -- a suitable form of comparator function. Maybe it's wrong, but it needs no conflict and it does produce avoidance without any need to suppose any alternative that is to be approached.
Now choice, that's a separate problem. How you make the choice of whether to walk, take your bike, or drive or take the bus is an issue. If "PCT" doesn't have a mechanism for choice, it certainly should. But is there a conflict to be resolved even here? All of the possible choices will serve to get you from A to B, but with different levels of convenience, expense, and travel time. I would, say, happily walk if someone had stolen my bike, the car wouldn't start, and I had no money for a bus ticket. I'm not avoiding walking because today I think it might be more fun to bike, which I haven't done for quite a while, and I don't see any conflict between the approach to the different possibilities, though I grant one could argue the point.
So I think we're agreed that some function resolves a conflict (one system wants to use the eyes and limbs to control a desired quantity of nuts and berries, another system wants to use the same means to control "no bears present"), and that Jeff's 'choice' function has problems. What do we put in its place as a corrective?
No we aren't agreed that a choice function must resolve a conflict. It may, but it need not.
When you think of what is available within "as proposed by Powers" version of PCT (which seems to be different from the Authorized Version), you have the associative memory proposal for the construction of reference values from past memories of what worked and what didn't work. Both the address into the memory and the output from the memory are likely to be vectors -- the address must be, and I think that is likely also to be true for the output, because we are talking about the context of the perception that was stored for later retrieval as a reference value. That context is a collection of perceptions of different aspects of the environment at that time, and the references provided as output from the memory presumably control as many perceptions as is necessary to construct a sufficiently similar context for the wanted action. Some contexts worked with one way of controlling a perception, some with another. That's the choice that the associative memory implements.
The berry-picker has picked berries in a context that does not include bears, and has avoided bears when bears are perceived to be nearby. If the berry picker is picking berries and a bear appears, the context has changed, and the associative memory is likely to provide reference values that would change it back to the right context for berry-picking (run away, if possible to another berry patch).
I don't think choice is a significant problem for PCT, even without using the polyflop. I suspect that we will eventually find that choice is implemented neither way, but in some way we have not yet considered. But that's for the future. For now, we have possibilities that seem to be functional, one with and one without modifications of the Powers version of HPCT.
Martin