00100	19 Jan, 1971
00200	
00300	Comments on "A Pred. Calc. Based Adaptive System" by David A. Plaisted
00400	
00500	1. No references - Has he read any literature?
00600	
00700	2. The notion of self-modifying is unclear, because one can always
00800	regard the modifiable part as data.
00900	
01000	3. A system which works by maximizing reward has to learn quite a bit
01100	before it can have goals.  It is better to start with a goal-achieving
01200	system.
01300	
01400	4. In general, this paper does not recognise various decompositions
01500	of the AI problem that have already been found.
01600	
01700	5. How specific is the specific system?  Is there a program?
01800	
01900	6. →Are expressions in the environment or the things they denote?
02000	
02100	7. There is no general answer to the questions raised in the section on
02200	maximizing reward.  Let some arbitrary answer to the questions be
02300	taken.  Then you still have to express in the language of the system
02400	what is being maximized, and this will be quite difficult in the cases
02500	that involve time.
02600	
02700	8. The idea that the result of evaluation is to reward or punish
02800	certain simple or composite actions is assumed without comment.
02900	I doubt that the useful kinds of learning will be conveniently
03000	described in these terms.
03100	
03200	9. I don't understand the examples on p. 6.
03300	
03400	10. Oho. 'u=P(x1,x2,...,xn)' really means  'u="P(x1,...,xn)"'.
03500	The expression on the right is not evaluated.
03600	
03700	11. The notation here can be done with LISP car-cdr chains or, more
03800	abstractly, using the notion of abstract syntax.
03900	
04000	12. "PAT" stands for "pattern"?
04100	
04200	13. The intent of p. 7 seems to be to provide a notation for expressing
04300	generalisations of collections of expressions.  The exact intent isn't
04400	clear, and examples are required for understanding.
04500	
04600	14. Damn it, Plaisted; How can you ask anyone to follow
04700	your "stream of consciousness" style of exposition.  After you wrote
04800	this, you should have revised it, incorporating afterthoughts with the 
04900	main concepts.
05000	
05100	15. EQPAT2 and friends are two involved to follow without examples.
05200	16. There may be good stuff in this induction on the form of expressions;
05300	at least, it's better than sequence extrapolation in that a real induction
05400	problem is more likely to be somewhere near this form.  The scheme needs
05500	to be tested on real generalisations, however.  The 
05600	association of the method with the goodness and badness of situations
05700	is an irrelevance.
05800	
05900	17. A problem with curiosity is disciplining it.  The curious
06000	system may spend the rest of its life studying the pattern of the
06100	plaster on the wall.
06200	
06300	18. An action is regarded as being done or not.  What about
06400	evaluating a parameter of an action, e.g. the distance to move
06500	the arm in reaching for the block.
06600	
06700	19. The introduction of probabilities is certainly an unnecessary 
06800	complication given the primitive state of the formalism.
06900	
07000	20. Separate the problem as follows:
07100	
07200		a. Separate learning what is good from goal achieving.
07300		b. Work out what a state of mind of the system is.
07400		c. Solve the problem of being told before the problem
07500	of learning.
07600		d. Delete all excrescences such as probability.
07700	
07800	21. I agree with the remark on p.25 about the potential advantages
07900	of parallel, argument centered programs.
08000	
08100	22. I suppose p26. asserts that when the system sees that P(x,y)
08200	is true, it is supposed to add Q(x,y) to its working set of
08300	assertions.  The inadequate general description leaves me still in
08400	the air.
08500	
08600	23. I wish I understood the example on p.27 and 28, because it
08700	looks like there might be a program there.
08800	
08900	24. The relation of your system is very hard to fathom, because
09000	truth in logic has nothing to do with proofs.  Therefore, the
09100	modification you want is ill-defined.
09200