First sketch of a rough outline for a paper entitled

Scientific Discovery by Computer




1. The hypotheses

Choice of domain is very important. 
  It must be knowledge-rich (vs, e.g., propositional logic), 
  have a strong theory about it (vs, e.g., most parts of most soft sciences), 
  be in a state such that the top experts are much better than the average practitioner 
	(and the experts are willing to talk!) 
	(or, at least where the task is performed by all but rarely performed well; e.g., math),

It suffices to consider the problem of discovery in science as a form of heuristic search.
  Supporting quote from Poincare'
  Expertise is demonstrated by doing the right thing at the right time; so heuristics
	must exist for guiding the selection of what to consider next, and what to do to it.
  The discoveries in a given field often will fall quite naturally into a few syntactic
	categories; thus the use of `templates' is sufficiently sophisticated.
        (note: this is true at 2 levels for AM: each concept follows a particular
	`concept template', and each conjecture must be of one of a very few
	formats or `conjec templates'. The latter seems closest to Meta-Dendral)

2. The experiments

  Meta-Dendral
	What it is (intent, design, representation and control structure)
	What it's done (heuristics present; best rules produced, a feelinn for its
		behavior both ypically and in extremely good/bad situations.)
	How this embodies -- and tests -- the hypotheses of section 1.
	The limitations of Meta-Dendral, and how that modifies the hypotheses.

  AM
	What it is (intent, design, representation and control structure)
	What it's done (initial concepts, heuristics; final concepts, conjecs).
	How this embodies -- and tests -- the hypotheses of section 1.
	The limitations of AM, and how that modifies the hypotheses.


3. Conclusions

  Summary of the previous section: a synthesis of how the results bore on the hyps.
  A more sophisticated model of scientific theory formation and scientific discovery
	by computer (derived directly by `debugging' the initial hypotheses).
  Suggestions for future work (on AM, MD, and on new topics) to test these new ideas.