I just noticed that while I annotated italics in the later sections, the first sections have un-annotated italics. I'm fine with italics, but we should be consistent and not over-use them.

I'm not sure whether we are allowed to have page numbers, but currently only the first page is numbered. We should either number all (if allowed) or remove the number on the first page.

Here is an annotated pdf with suggestions for improvements. Note that this pdf is not the latest version available (we move too fast for that 😜), so some sections that I know have been worked on are not annotated. These sections will need to be reviewed separately (please mention those below so we won't forget them).

Section 3 and 4 currently discuss mostly the same things; these need to be merged. Therefore neither section is annotated (also because section 4 is my own, which somebody else should check).

When you fix or comment on an annotation, please be as explicit as possible as to which annotation we are talking about so I can keep the pdf updated appropriately.

main_annotated.pdf

Just to track things we need to make consistent:

• labelled -> labeled and modelled -> modeled, basically all lled should be verified to be correct (#37 - #46)
• input output and I/O to input/output (#61)
• DFAs instead of DFA's (#37 - #46)
• behaviour -> behavior, basically all iour should be verified to be correct (#37 - #46)
• Write out small numbers (#37 - #46)
• In some places $\row(s)$ is used and in the example run it's row$(s)$ (#47)
• Use \mq macro in L* part (#47)
• Consistent use of $L^*$, so not L*, or $L*$ etc. (#47)
• Refer to everything as sections, not chapters. (#63)
• "membership queries" vs. "$\mq(w)$ queries". I (Skip) vote for the former. (#63)
• When citing, it should be "Howaretal." when referring to a publication by three or more people. Not "et. al.", and not "et al" (#63)
• Equivalence classes should be w.r.t. the hypothesis, not the target. (#63)
• Citation methods. Skip has seen the following ways of citing throughout the document (some may be compatible):
  1. "text as found in \cite{author year}"
  2. "According to author, text \cite{author year}" or "authors note that..."
  3. "text \cite{author year}"
  4. "In year, author text \cite{author year}"
• Non-breakable space before each citation and reference. (#65)
• ..., and ... to ... and .... I know this is a disputed style change (at least @StefanBoodt and @pvanagtmaal disagree with me), but most of the article has already been changed to the latter so I think we should straighten this out. Alternatively, if someone is really passionate about having their commas, we can also go over the article as a whole and put them back.
• mealy to Mealy (#37 - #46 )
• Make better use of SUL instead of system under learning (#61)
• Fully write "do not", "does not", "cannot", etc instead of "don't" (grep -R "n't" *) (#37 - #46)
• Remove all instances of "we" (#61)
• Remove all verbal shorthands (we'll, don't, doesn't, anything with a letter on both sides of ' if it isn't possessive form). (#61)
• Consistent use of learner vs Learner (#61)
• Consistent use of teacher vs Teacher (#61)
• Consistent camelcasing of (sub)sections (#61)

Please edit this post to add more.

Adding (base) to each list entry is redundant, as the sentence describing the list already mentions that the algorithms are base algorithms.

Mandatory:

• #32 Thoroughly read variants, read Chow's method
• #40 Write abstract (Piet)
• #54 Write conclusion (Piet)
• #42 Restore Libalf section (Arjan)
• #60 Restore applications section (Piet)
• #41 H2 L* pseudocode (Gerlof)
• #50 Chow pseudocode
• Move an example of observation table to top of 2 to explain structure (Sander)
• #43 Observation pack / TTT (Skip)
• Transforming mealy machines, find reference for own conclusion (Sander)
• #49 Expand limiting membership with practice (Gerlof)
• #44 move parameters in 3.4 to its own section and talk about learning other automatons
• #35 Consistency fixes
• Sicco feedback: L*_M show example observation table (Sander)
• Sicco feedback: L*_M success guaranteed? (Sander)
• cite reason why random sampling is used

Other additions:

• #30, #57 H3.1 Random sampling

Formatting issues:

• #26 Italics
• #25 Page numbers
• #33 Cite with page numbers (not required)
• Observation table mark equivalence classes (low priority)

In Table 1c and 1d , the answer of row(bbb) under column b is wrong. Instead of a '1' , it should be a '0' since the word "bbbb" is not accepted by the target.
The answer of row(bbba) under column b is wroing. Instead of a '1' , it should be a '0' since the word "bbbab" is not accepted by the target.
The answer of row(bba) under column b is wroing. Instead of a '0' , it should be a '1' since the word "bbab" is accepted by the target.

To bring back some removed points in #34:

• #34 (comment)
• #34 (comment)
• #34 (comment)
• #34 (comment) @gfokkema restart SUL

This text from section 3.4 should be moved or removed and described in this section:

Another obstacle to overcome is that of parameters used in real world applications. Think,
for example, about increasing sequence numbers in communication protocols. This is still a huge
challenge to resolve [37] and for now, besides the creation of prototypical solutions [4, 35, 16],
application-specific solutions have to be applied [37].

Sicco also specifically mentioned learning other automatons, such as register automata.
Doesn't have to be very elaborate, but explaining in general about mapping, reactive systems and register automata and such should be enough, using @Unia his fragments linked above.

Other leftovers from #34:

• #34 (comment) As stated on telegram, fixed already.
• #34 (comment) @gfokkema add ref to TTT
• #34 (comment)

I think the "Approximating EQUIV (M) queries with MEMBER(w) queries" subsection should contain a short subsection on random sampling.

#28 (comment)

Create new section 3.5, about:
Reactive systems
Mapping alphabet
Mapping
Register automata

This text from section 3.4 should be moved or removed and described in this section:

Another obstacle to overcome is that of parameters used in real world applications. Think,
for example, about increasing sequence numbers in communication protocols. This is still a huge
challenge to resolve [37] and for now, besides the creation of prototypical solutions [4, 35, 16],
application-specific solutions have to be applied [37].

@Gerlof: add this somewhere.
Restart SUL
Reference TTT

I just remembered that during our last meeting, Sicco mentioned that he wanted the L* as an algorithm in the article.

All the above problems might give the impression that active state machine learning is not yet applicable to real world applications. While it is true that practical application is not yet com- plete, great progress has been made and great things have already been achieved. The last chapter serves to highlight some real world applications that illustrates these facts.

Should be either before the chapter on applications, or the applications intro should be edited to contain the above.

Using \cite[p. xx]{reference}