Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!usc!cs.utexas.edu!uunet!mcvax!ukc!stl!dsr
From: dsr@stl.stc.co.uk (David Riches)
Newsgroups: comp.ai
Subject: Re: Adaptive vs. intelligent (was Re: "Intelligence")
Message-ID: <1556@stl.stc.co.uk>
Date: 27 Jun 89 14:53:06 GMT
References: <1533@stl.stc.co.uk> <1597@infmx.UUCP>
Sender: news@stl.stc.co.uk
Reply-To: "David Riches" <dsr@stl.stc.co.uk>
Organization: STC Technology Limited, London Road, Harlow, Essex, UK
Lines: 187

This contains the reply from Pete Totterdell who introduced the
Taxonomy of Adaptive Levels into the Project.

Here are some comments for Brian:

I am not sure how much information Dave gave you. If it was just the
taxonomy without any explanation then I can understand why you had
problems with it! Unfortunately I cant reproduce the whole argument
here. If I did I think you would probably see where all these levels
have come from, although you wouldn't necessarily agree with them.
Here are some specific points.

In article <1597@infmx.UUCP> briand@infmx.UUCP (brian donat) writes:
|
||---------------------------------------------------------------------
||Levels	Computers	Evolution	Features
|
||0.5	Tailorable/	-		Deferred selection
||	Adaptable
|
|OK, so here we have a system which (by the term tailorable), is assumed
|to solve problems based upon what functionaly is a rigid switch, for a
|decision maker.  Each activation of the switch gives a defined output, 
|whether the switch is activated by a single or multiple inputs.
|
||1	Adaptive	Tropism/	Apparent Learning
||			reflexes	(i.e. fully determined by design)
||					Discrimination
|
|By the term 'Apparent Learning' it is assumed that it realy doesn't fit
|the definition although it appears to.    So perhaps we have here a 
|system which is only slightly more evolved than the rudimentary switch.
|In such a system, it might be assumed that it alters its composition so
|that as it experiences an event, it remembers the event by storing something
|or by altering something so that the next time the equivalent event occurs,
|its response is 'tuned' and it need not waste time analyzing and just does
|the appropriate thing.
|
|FINE.  OK.    Again it appears that the programmer controls outcomes
|although this time, the inputs and sequencing of responses need not be
|rigid.   Inputs (environment) begin to influence the system's character.
|

Apparent Learning. There is no remembering going on here. Hill climbing
(either biological or AI style) is a good example. The system finds the
best solution to a problem (and gives the illusion of learning) simply
by having fixed responses which are controlled by the environment. A
plant growing towards light is an example.

|
||2	Self-Regulating	Operant		Learning; Varied responses
||			conditioning	selected for different situations;
||					Evaluation by trial and error
|
|Another evolution?    What could this be?     Real learning?   
|
|I'm sorry, but I'm lost here.   This seems to be something similar to
|level 1, except that you're saying, that it actually 'tries' multiple
|inputs sequentially and selects one for subsequent processing based 
|upon some programmer defined criteria.
|
|I don't see learning here.   Learning would mean that the system 
|auto-generates new code (it's own response) to go beyond the 
|programmer's built in restrictions, tests the new code and then, 
|judges which code handled the input situation better and keeps it for
|subsequent 'reference'.   I say reference here, because this leaves
|the door open for the system to reject implementation of the same response,
|if causal values should weigh it's decison to do so.   Learning then 
|becomes the machine's confidence in the same solution.   
|
|I believe what we have so far, is three examples of tailored adaptation
|with varying complexities.
|

At this level, where we have genuine learning, the system has a
number of different responses to a given stimulus and it learns (and
sometimes remembers) which is the best. But the best response may
change as circumstances change.

|
||3	Self-Mediating	Internal	Planning, Problem Solving;
||			evaluation	rule-mediated representation;
||					Initial evaluation internal to
||					system
|
|
|What is planning without the ability to recognize a problem?   What is
|a problem to a system which can not 'generate' problems?   When you 
|say 'problem solving', each of your earlier levels seem to have as
|much to do with this same type of problem solving as this level.    
|The system is solving your problem, ultimately, based on your rules!
|
|The problem is given by you.  You expect it to give a solution which you
|have already defined.   This system is no less 'programmer rigid' than 
|those you site in the first two levels of adaptability.
|
|Intelligence must be autonomous to the point that it observes less formal
|rules than 'tailored rules'.   If you really had a system which was
|at an adaptive level such that it could 'plan' and 'solve problems' based
|on planning, you'd have a system that should be able to 'define and solve
|its OWN problems'. 
|
|Perhaps my perspective is wrong, but it really does seem like you're 
|just building a more complex machine which is really still operating on
|the same adaptive principles (limitation to programmer definition) which
|you have already defined in level 0.5 as 'tailored'.

The planning referred to at this level is meant to cover the sort
of insightful behaviour that you refer to when you talk of generating
new problems and solutions. An example is the ape which learns that
to reach food it needs to reconstruct the problem and use the short
stick to reach the big stick which is out of reach and then use the
big stick to reach the food. However, I take the point about
problem solving. All levels are problem solving, although I think that
AI uses problem solving in the more restricted sense of this particular
level.

||4	Self-Modifying	Abstraction	Evaluating the evaluation;
||					Generalisation; Meta-knowledge
|
|This doesn't help much.   Looking back on results of programmer defined
|tests is just another conditional in your switch which amounts to another
|programmer defined test and thus is also tailored.
|
|
|Perhaps these suffice as 'primitives' for a single adaptive level
|which is defined by variations which are tailored.
|
|
|Your level 2 'primitives' might better allow introduction of a device for
|recognizing the 'need' for a response such that the recognition of the 
|need and the response to it are both auto-generated, unless the response
|is already conditioned; in other words, two part problem resolution where
|problems are auto-defined as a response in themselves and actions are
|not only taken from a known selection of givens, but are created as 
|required and the givens become mallable over time.
|
|The machine should create it's own problems and resolutions and it should 
|do this for all situations which first threaten it's survival, prevent 
|it's own self-demise and for those situations whereby it enhances it's 
|own survival and then, for general and logical resolutions.
|
|
|If you evolve an adaptive system which can guarantee its own survival from
|its own recognition of problems and its own resolutions to those problems,
|you'll have obtained a level two adaptive system.
|
|
|Level 3 would inlcude problem solving of the type which goes beyond 
|mere survival and defines and resolves problems which are in the realm 
|which are common to we humans alone (in terms of life on this planet).
|
|
|Everything must start somewhere, but I believe your aim's a bit low.
|
|

My main departure from your argument is that I get very worried when I
see terms like "auto-generate" which often suggests that people are
looking for something from nothing in order to give a sense of free
will, and to get away from pre-determination. Biologists seem happy
with the idea that rats are learning when they learn to run a maze
and yet we do not need to invoke explanations involving auto generation
in order to explain this type of behaviour. And of course even in the
situation where you are generating new problems and new solutions, the
antecedents will have been pre-determined by the programmer. What the
programmer doesn't control however is the course of the interaction
between system and environment.

I also felt that your third level, about going beyond survival, would
shift the ground somewhat because it describes the level in terms of
a type of problem which the other levels don't. And in many ways it
is simply a bi-product of the fact that all the levels are about
trying to bring the environment under control. However, it might
be the basis for another equally valid taxonomy which looks at problem
types.

Hope this makes sense .. Peter (Peter Totterdell pc1pat@uk.ac.shef.ibm)


   Dave Riches
   PSS:    dsr@stl.stc.co.uk
   ARPA:   dsr%stl.stc.co.uk@earn-relay.ac.uk
   Smail:  Software Design Centre, (Dept. 103, T2 West), 
	   STC Technology Ltd., London Road,
	   Harlow, Essex. CM17 9NA.  England
   Phone:  +44 (0)279-29531 x2496