00100	COMMENT ⊗   VALID 00007 PAGES
00200	C REC  PAGE   DESCRIPTION
00300	C00001 00001
00400	C00002 00002	This file is a compendium of ideas about Parry.  NEWPAR[NEW,BLF].
00500	C00008 00003	More ideas
00600	C00014 00004	
00700	C00016 00005	ways to go beyond the current internal workings:
00800	C00021 00006	        THIS IS FOR NEW PARRY IMPROVEMENTS
00900	C00025 00007	Comments on improvements to the theory:
01000	C00026 ENDMK
01100	C⊗;
     

00100	This file is a compendium of ideas about Parry.  NEWPAR[NEW,BLF].
00200	IDEAS FOR THE FRONT END HAVE MOVED TO NEWPAR[HMF,RCP].
00300	
00400		The purpose of  this file is  to give us new  direction about
00500	where we want  to go with parry and its associated research problems.
00600	The file contains criticisms  of old parry,  ideas about a new  parry
00700	which we think  are possible, and dreams about the  way it all should
00800	be  done in  an ideal world.   Please feel  free to  comment on other
00900	comments.
01000	
01100	
01200	
01300	parry cant  pick  up answers  to his  own questions  --  there is  no
01400	mechanism that tells him when  he has asked a question so he can look
01500	for an answer
01600	
01700	parry cant refer to his own thinking  patterns -- if you ask him  why
01800	he thought  something (e.g., why  do you think  i am  incompetent) he
01900	doesnt know.   Also, if you ask what kind  of general rule he uses to
02000	discover a fact, he doesnt know.
02100	
02200	KEN. BOTH OF THE ABOVE MEAN THAT PARRY NEEDS A "SECOND LEVEL" IN HIS
02300	SYSTEM WHEREBY HE CAN LOOK AT, MONITOR, REPORT ON, AND
02400	REACT RECURSIVLEY WITH HIS OWN STATES IN ORDER TO CHANGE HIMSELF
02500	THROUGH SELF-INSTRUCTIONS.
02600	all of the memory data bases are difficult to add to.   This includes
02700	the  main memory  of  unchanging  facts, the  beliefs,  the rules  of
02800	inference,  the affect changes,  the intentions.   In addittion, when
02900	changes are made, it is not clear that it  wont have disasterous side
03000	effects on other data in the data base.
03100	
03200	I cant think of a way  to verify individual processes.  The way it is
03300	done now is to put  it in the system parry  and see if any bugs  come
03400	along.
03500	
03600	KEN.. ONE WAY TO VERIFY INDIVIDUAL PROCESSES IS BY SENSITIVITY
03700	ANALYSIS. FOR EXAMPLE, THE PARANOID MODE IS SWITCHED ON WHEN
03800	SHAME RISES OVER A THRESHOLD. IF WE KEPT SHAME STUCK AT SAY 2,
03900	THEN JUDGES SHOULD NOT DEEM HIM PARANOID SINCE ONLY NORMAL
04000	PROCESSES ARE BEING ACTIVATED. WE COULD EVEN TURN ON THE MODE
04100	AND TURN IT OFF DURING A LONG INTERVIEW SO THE RATINGS
04200	WOULD (HOPEFULLY) LOKOK LIKE THIS 00000006787767530012000.
04300	The static data base is  not touched by the belief system.   It would
04400	be nice if the static base could be included in the belief system.
04500	
04600	The entire motivation should come from a top-level affect routine.
04700	
04800	The  task of participating in  an interview and  the knowledge needed
04900	for such participation is currently implicitly spread over the entire
05000	system. It  would be nice to  separate out that knowledge  so maybe a
05100	new  task could  be undertaken someday  by just  adding a  new set of
05200	facts.
05300	
05400	The interaction  between the various  pieces of  the memory is  still
05500	vague and hard  to see clearly.  It would  be nice if the interaction
05600	could be made clear and visible  and defined so that any big  changes
05700	in one part  could be made while  clearly knowing what the  effect on
05800	the rest would be.
05900	
06000	PARRY STILL  HAS TOO MUCH OF OLD PARRY. THE WHOLE THING SHOULD BE
06100	REWRITTEN IN A NICE CLEAN WAY. NO KEN..
06200	Currently  there  is   no  way  of  testing  whether  the  memory  is
06300	functioning correctly  or even  pieces of  it are  correct except  by
06400	putting parry on the system for  20 dialogs and looking for errors in
06500	dialogs.
     

00100	More ideas
00200	
00300	Currently the inference system can only be called once per io pair at
00400	a specified  time.  We should be able to  call it any time, interrupt
00500	it  when important  results  are  found,  resume,  have  it  work  on
00600	different  kinds of  tasks, such  as figuring  out a  good action  to
00700	satisfy an intention.
00800	KEN.. IDEALLY THIS SHOULD GON ON WHILE PARRY IS WAITING FOR INPUT
00900	BUT I GATHER THERE ARE SOME HARDWARE PROBLEMS TO THIS.
01000	
01100	It would  be nice to be able to add more history data without as much
01200	trouble as at  present.  For  instance, tell some  system that a  new
01300	question  can be  asked,  tell it  the  facts involved,  tell it  the
01400	answers to the questions, the importance of the topic, how it relates
01500	to other main topics.
01600	
01700	All of the  kludgy mechanisms that  parry uses to keep  a converstion
01800	going -- story lists, flare topic lists, GO_ON anaphora -- as well as
01900	the  mechanisms  for  discovering  facts  about  the  style   of  the
02000	conversation -- conversation list, topics,  topic counts -- should be
02100	put into  the same style of program and data rules as the rest of the
02200	system.  i.e., no more special case kludges.
02300	
02400	parry cannot remember things from one dialog to the next.  This might
02500	be useful someday.
02600	
02700	parry needs to be able  to examine his past life and comment upon it,
02800	and have feelings about it.
02900	
03000	perhaps parry could go thru one  interview a day where he  remembered
03100	what was going  on and had  feelings about it.   there could  be many
03200	history files starting at the same place and progressing from one day
03300	to the next. each history file would have beliefs and  feelings about
03400	the history to the present.  one history file per interviewer.
03500	
03600	the  belief  system  should  be extended  to  handle  thoughts  about
03700	intentions, actions, affects, plans.
03800	
03900	the beliefs themselves could be generalized in form -- e.g., a person
04000	is hostile, the person is the doctor.  then  generalize the inference
04100	rules -- e.g., if a person is hostile,...
04200	
04300	the  calling structures  of each  nice piece  of mechanism  should be
04400	general to allow calling an intention  and then thinking about it  or
04500	emoting about it.
04600	
04700	too much of current parry's anger and  fear are expressed randomly in
04800	his  hostile canned replies.   it would be nice  to really categorize
04900	the type of emotion which the  memory wants to display and then  have
05000	fine lists for the specific emotion and its appropriate strength.
05100	
05200	KEN.. YES. PARRY IS STILL TOO TERSE, ANGRY, AND UNFRIENDLY.
05300	yes and  no don't  work in  parry.  it  may be  just an  extension of
05400	answering  a question, or answering a questing  may be just part of a
05500	larger problem of a dialog input which depends heavily  on the syntax
05600	and  semantics of  either  he previous  output,  or the  interviewers
05700	previous input.
05800	
05900	three properties of a good  understanding system: generality -  could
06000	be extended to more  topic areas and deeper understanding  learning -
06100	it  can  extend  itself  in  some  way  dynamically  cleanliness  and
06200	complexity - it is complex enough to be instructive, and clean enough
06300	to be learned from.
06400	
06500	three  effects  of emotions  -  arousal,  discomfort, and  confidence
06600	(anxiety)
06700	
06800	KEN.. In next go-round Parry needs affect of contempt-disgust to
06900	generate contemptuous expressions
07000	
07100	what is an action?  is it just an output expression? or could it also
07200	be concluding a belief, modifying an affect, incrementing an intention,
07300	testing a belief, i.e., any action at all in the system?
07400	
07500	What is the interaction of conation and cognition?  When does cognition
07600	call conation for subgoals and subactions?  What part of cognition is
07700	an action itself and needs to be in conation?
07800	
     

00100	
00200	WE NEED SOME WAY OF TAKING THE IYOUME AND GENERATING APPROPRIATE
00300	RESPONSES WITHOUT HAVING A SEPARATE SF AND SET OF RESPONSES FOR EACH.
00400	
00500	
00600	REPEATED HOSTILE REMARKS OR INSULTS OR THREATS SHOULD BE ADDITIVE,
00700	  RATHER THAN JUST ONE-TIME.
00800	
00900	TRANSFER THE APPRO STUFF FROM DOINTENT INTO INF.
01000	
01100	MAY WANT TO LET ANGERFEARMODE TAKE OVER ON LOWER VALUES WHEN DDHARM, ETC,
01200	  AND HIGHER VALUES WHEN NO BAD INFS ABOUT DOCTOR.
01300	
01400	SOME GENERAL WAYS OF STICKING COMMENTS ABOUT DOCTOR AND INTERVIEW IN THE
01500	  INTERVIEW, AND WHAT THESE COMMENTS WOULD BE.
01600	
01700	
01800	Have a system of explicit beliefs for what happens in interviews.  Then
01900	use these beliefs when doing inferences and intentions.
02000	
02100	Don't use the numerical characteristic of beliefs so much.  I.e., instead
02200	of adding 2 to ddharm for a consequent, maybe the level of ddharm is
02300	an additional antecedent for theorems, and then the consequent is
02400	to set the belief to a specific level.  But then what happened to the
02500	idea of a threshhold?
02600	
02700	
02800	
02900	
03000	
03100	
03200	
03300	
03400	
     

00100	ways to go beyond the current internal workings:
00200	
00300	1.  all of the inference rules are fixed in advance and work only with specific
00400	beliefs.  would be more elegant to allow rules of the form:
00500	  i believe(x) ∧ he says (¬x) →→ he lies
00600	
00700	or to allow rules of inference from the input (sigh)
00800	  ∀x xεA ⊃ xεB
00900	
01000	2.  affects and intentions should influence recognition and the inferencer (how?)
01100	
01200	
01300	3.  conflicting evidence for beliefs -- maybe inhibit each other, and/or partial
01400	cancellation 
01500	side effects of discomfort or uncertainty
01600	
01700	
01800	4.  compare the self to some standards of normalcy
01900	
02000	
02100	5.  dont want to make all possible inferences, just interesting ones
02200	
02300	
02400	6.  the driving force of parry -- some affect (or need) for self-preservation
02500	and/or self-maintenance.  from this could be an intention to interact, with
02600	subintentions (plans?) of how to interact in an interview.
02700	
02800	
02900	7.  reporting and thinking about some of the chain from basic affect thru
03000	intention and inference to action.
03100	
03200	
03300	8.  once a belief is true, is partial belief in negation possible (or even
03400	psychologically right?)
03500	
03600	
03700	9.  can part of a plan for an intention be to check the next input to see if
03800	the action was successful.
03900	
04000	
04100	10. planning could be the phase from intention to action to execution of action
04200	
04300	
04400	11. make a distinction between causative and resultative inferences?
04500	  types of causative: 
04600	    a) cause of situation x  b) cause of speakers belief that x obtains
04700	  enabling conditions:
04800	    a) intrinsic - powers of individual  b) extrinsic - environmental circumstances
04900	
05000	
05100	12. more theory into ordering and selection of intentions -- what intentions
05200	are there, what does that say about parrys past life, can the ordering be
05300	changed dynamically, how is parrys feeling about life reflected in his intentions
05400	and their selection, how are affect requirements reflected in intentions
05500	ex -- current intentions imply a withdrawal strategy
05600	
05700	
05800	13. predictions about the interviewers behavior, based upon a comparison to
05900	normalcy
06000	
06100	
06200	14. the difference between analysis of nl input and all the processes required
06300	to use and respond to the input
06400	
06500	
06600	15. the higher process that detects the potential humiliation and prevents it
06700	from happening -- what is this process, what else does it do, what are its goals
06800	
06900	
07000	16. how much of these processes are conscious -- particularly, how much of the
07100	paranoid mode humiliation detection and refusal are paranoids and parry
07200	conscious of.
07300	
07400	
07500	17. style info -- how about negations of affect words:  you are bad, you are
07600	not nice, you are not too bad, you are nice.
07700	
07800	
07900	18.  confidence level in pattern matchers input -- should be calculated by pm.
08000	
08100	
08200	19.  some rules can be stronger than others -- have strengths associated
08300	with the rule of inference, so that a little evidence for a strong rule
08400	does something like a lot of evidence for a weak rule.  This should be 
08500	compared with adding only partial evidence to the consequent when a rule
08600	of inference's antecedents are all true.
08700	
08800	
08900	
09000	
09100	
     

00100	        THIS IS FOR NEW PARRY IMPROVEMENTS
00200	PARRY SHOULD BE ABLE TO COMMENT ON
00300	ITS OWN STATES AND ACTIONS. FOR EXAMPLE, "I SAID
00400	THAT BECAUSE I THINK YOU ARE INCOMPETENT". 
00500	
00600	TO BE COMPLETE,
00700	 A THEORY SHOULD ACCOUNT FOR ALL KNOWN
00800	OBSERVATIONS AND EXPERIMENTS. THIS IS A STRONG REASON
00900	WHY PARRY SHOULD HAVE MECHANISMS FOR PHENOMENA OF GRANDEUR.
01000	
01100	TO BE CONSISTENT, A THEORY SHOULD PREDICT SAME RESULT
01200	NO MATTER HOW IT IS CALCULATED. (OBSCURE)
01300	
01400	TO BE A THEORY AT ALL, IT SHOULD PRODUCE MODELS. THE
01500	MODEL MAY NOT BE CORRECT BUT AT LEAST IT SHOULD EXIST.
01600	NOTICE THAT IN GRAVITATION THEORY, WHEN THE "GRAVITY" LAWS
01700	ARE SUPPRESSED, THE NON-GRAVITATIONL LAWS HOLD TO THE SPECIAL
01800	THEORY OF RELATIVITY. THUS IN PARRY, WHEN WE SUPPRESS THE
01900	
02000	WHAT MAKES A THEORY A VIABLE CANDIDATE IT DOES NOT VIOLATE
02100	ANY OF THE CRITERIA SET UP FOR AN ACCEPTABLE THEORY. THE CRITERIA
02200	ARE (1) COMPLETENESS (2) SELF-CONSISTENCY (3) APPROXIMATES
02300	NEWTON'S LAWS (AS LIMIT) (4) RELATIVISTIC.
02400	
02500	WHAT IS THE COUNTERPART OF THIS IN THEORIES OF PSYCHOPATHOLOGY.
02600	
02700	
02800	
02900	MAY 28 -- FROM VAUGHAN
03000	
03100	IT WOULD BE GOOD TO CHANGE PARRYS OUTPUT SO HE WOULD LOOK NEWER, SO THE
03200	VARIABLES WOULD NOT GIVE PEOPLE THE OLD PARRY IMPRESSION
03300	LIKE MAYBE WHAT PARRY IS THINKING, WHAT THE INTERVIEWER IS THINKING,
03400	WHAT PARRY WANTS.
03500	
03600	THE DIFFICULTY WITH COMPLEX SYSTEMS LIKE PARRY IS THERE IS NOT ENOUGH SPACE
03700	WITHIN MY HEAD TO VISUALIZE THE WHOLE EFFECT.
03800	IT WOULD BE NICER IF EACH FACTOR (A BELIEF, INTENTION, EMOTION, WHATEVER)
03900	COULD BE A SEPARATE MACHINE, WHICH WOULD ACT ON ITS OWN, SPECIFYING ITS
04000	STATES, INPUTS, OUTPUTS, AND THE RELATIONS BETWEEN THEM, AND WHAT WOULD
04100	TAKE YOU TO NEW STATES.
04200	THEN I COULD LOOK AT EACH PIECE SEPARATELY AND SEE IMPROVEMENTS TO BOTH
04300	NOTATION AND ACTUAL DATA.
04400	
04500	THE IDEA IS PULSE AND LEVEL LOGIC.  EACH MACHINE CALCULATES AN INTERESTING
04600	PIECE OF INFO.  THIS MAY BE AROUND FOR ONE PULSE, OR MAY BE STEADY STATE.
04700	
04800	
04900	
05000	
05100	
     

00100	Comments on improvements to the theory:
00200	
00300	
00400	Simulation of blame path, too, in addition to simulation of shame path.
00500	  Then compare the two, and choose the blame path.
00600	
00700	Careful with masochistic tendencies: they can be used to explain most
00800	  any emotional disorder.
00900	
01000	
01100	
01200	
01300	
01400	
01500	
01600