Aminet 18/aminetcdnumber181997.iso/Aminet/dev/misc/Istar.lha/Docs/Driving Istar

Text File | 1997-02-10 | 23.1 KB | 485 lines

DRIVING ISTAR

WHAT DO I NEED TO START IT?

Normally you don't need to bother about this lot as they are set up during
installation and Amiga startup.  But if something goes wrong then this list
might help.

#  assign KBTools: to some directory for best effect
#  assign IstarSys: to some directory for prefs files (and, in future,
other things)
#  iff.library (can run without it)
#  asl.library is needed for saving etc. (supplied with Amiga 1200)
#  Fonts needed depends now on IstarSys:IstarPrefs but the defaults are:
      #  Courier 11
      #  Garnet 16
(all supplied with standard 1200).

HOW DO I START IT?

This describes the normal start of Istar, from WorkBench.

(If you are operating with low memory, such as a base A1200, then refer to
the section below on 'Starting with Low Memory'.  If you start from CLI, a
few options are available in Version 1.04 onwards; see later.)

Note:  Bold is often used for concepts and jargon terms used in Istar when
they are first used.  Italics is often used for some action you should
take.  'Oops' is used to tell you what to do if things don't seem to be
going as expected.


1.  Double click on the Istar icon.

2.  You will be presented with a grey Control Screen part way up the view.
Initially this has 'Istar' on its title bar but that is soon obscured by a
horizontal strip and, at the left, the Initial Control Panel.

3.  Click the 'New' button, and two things will happen.  First another
screen, a large blank white one appears.  This is the Easel on which you
draw your KB (knowledge base).  And, on the Control Screen, another panel
appears with columns of buttons and two list gadgets.  (If you are
operating with limited memory, the standard Easel size (30 by 24 inches)
might be too large, and you will have to alter the Prefs before clicking
the New button.)

Here are a few instructions to get you started ...

4.  Move the mouse pointer up over the large white Easel then press left
mouse button (LMB) and keep it down.  A box appears that moves with the
mouse.  Move the mouse around.  Then release LMB.  (Oops:  No box appears.
Have you clicked rather than press, drag, release?  Do not click it;
clicking does something else.)

5.  Draw another box, in the same way, to the right of first one.

6.  Move the mouse pointer tip over the right hand edge of the left-hand
box, and press LMB while over the edge.  You will see a short line appear
from your mouse pointer to a point one quarter the way into the box.  This
line follows the mouse as you move it.  Move mouse around until line is
over the other (right-most) box and release LMB.  (The left-hand box is
called the antecedent of the right, and the right is the consequent of the
left.)  (Oops:  No line appears, but the box moves?  You had the mouse
pointer too far into the box, not exactly over the right-hand edge.  Oops:
A new box appears?  You had the mouse pointer too far outside the box, not
exactly over the right-hand edge.  In either case hit Escape key; see 7
below.)

6a.  (If you wish, you can draw the line in the reverse direction: from the
left-hand edge of the right-most box leftwards to the left-most box.)

6b.  If you want to insert a bend in your line, hit Space bar when end of
line is where you want the bend.

7.  (To escape from any operation you have started but don't want, hit the
Escape key.  Then release the mouse button.)

That's the basic mechanism, and the whole knowledge base can be drawn like
that.  Not too bad, is it?


GOING FURTHER

8.  Various modifications of the diagram are possible:

      a)  Move a box by pressing LMB over the centre of the box.  Dotted
      skeleton of the box and all its connecting links appears.  Move mouse
      and this lot follows.  Release in new position.  (Note that boxes can
      overlap, though it is usually not wise to do so; overlapping is
      usually reserved to collections of boxes that are closely related
      semantically.)  (Oops:  No dotted skeleton appears.  If a new line
      appears then you have pressed the LMB while over a right- or left-
      hand edge of the box, not near its middle.  Hit Escape and try
      again.)

      b)  Bend a drawn link by pressing LMB over a link.  Dotted bent line
      follows mouse to new position; release.  Pressing LMB over existing
      bend will just move the bend to new position.  (Oops:  line doesn't
      bend, but a box appears instead?  You were not near enough to the
      line when you pressed LMB.  Hit Escape and try again.)

      c)  (Other modifications are possible e.g. to redirect a link to
      another box, or to delete a box, delete a link, delete a bend from a
      link:  see below.)

9a.  Entering detail, e.g. Name and Meaning.  Click the Right Mouse Button
(RMB) over centre of box.  (Oops:  If you get a dotted box outline as in
8(a) then you haven't clicked fast enough; hit escape key, release MB and
try again.  You'll soon get the hang of it.)  Up comes the Control Screen,
showing an Attribute Detail panel with around fifty gadgets.  Don't panic:
you need only the top two: Label and Meaning; ignore the others for now.
(Oops:  If a small panel comes up with half a dozen buttons - the Action
panel - then you have clicked the wrong MB; as we shall see below, this
comes up when you click the LMB rather than the RMB.  Hit 'Done' button
with the LMB to send it away and try again.)

9b.  Along top of panel are two string gadgets which allow you to enter a
name for the box and a longer meaning if you wish.  For now, type something
in Label string, press Return and carry on typing in Meaning string - type
anything that you like.  Then click the OK button.  The Control Screen
returns and your box should display the new name.  (Oops:  If you try to
type and nothing enters either string then click LMB over the string you
wish to type in.)

10.  Showing Meaning.  On the Easel, move the mouse over the box (do not
press any mouse button).  The Meaning text should appear in the horizontal
window in the top of the Control Screen.  Move mouse away, and it
disappears.  In this way, even when the box label text is too small to read
you can easily see what a box means.

10a.  (Oops:  Nothing happens?  Probably the Easel has not been reactivated
after hitting the OK button.  Only one panel or easel is active at any
time.  If you see the border of a panel is pink rather than grey then that
panel is active, rather than the Easel.  You can activate the Easel in
several ways.  One is to hit the 'See Easel' button on the control panel
(not available on earliest versions of Istar).  Another is to move over the
centre of a box on the Easel, press LMB in an attempt to move it - see 8a
above - and release or hit Escape key.  Anything using LMB over Easel will
activate it.)

11.  More antecedents.  Draw another couple of boxes (see 4 above) near the
left-hand box (above or below it) and link these to the right-hand box as
in 6 above.  The right-most box now has three or more antecedents.  Best
also to name them with different names as in 8 above.

12.  Running the inference net.  Now click with the Left Mouse Button over
the centre of the right-hand box.  Up comes a small panel at the top of the
Control Screen containing a few buttons, the Attribute Action panel.  The
two buttons on the left are: Reset and Infer, which allow you to run the
KB.  Hit Reset, then hit Infer.  Up comes a User Question panel, with text
at the top showing the name/label of one of the antecedent attributes, a
large central Value gadget, a slider, and an large OK gadget to right, with
a few others.  Enter a Value by sliding the slider over the the right using
the LMB and hit OK with LMB.  Then a similar panel should appear for the
next attribute.  And the next.  If you have three antecedent (left-hand)
boxes feeding into the consequent (right-hand) one you should get three
such User Question panels.  (Oops:  No panels come up.  Have you hit Reset
then Infer?  Oops:  Only one Question panel comes up.  Maybe you have the
Action panel for one of the antecedent attributes?  A glance at its title
will give the name/label of the attribute for which it is the Action
panel.)

12a.  (The meaning of the values given by the sliders will become clear
when you start the chapter Using Istar.)

13.  Seeing the result.  When the question sequence has finished, the
Attribute Action panel is still there.  Hit the Data button, and you are
taken to the Attribute Details panel.  The result should be shown in the
slider gadget on the second row in the middle - and it should show
somewhere over to its right end if you slid all the antecedent sliders over
to the right.

14.  Control Screen.  You can drag the Control Screen up and down using the
thin white line along its top, just above the strip where the meaning texts
appear.  Use the tip of the mouse pointer and LMB to do this.  Half way up
is not a very convenient place, so you can park it by dragging the screen
down to reveal more of the Easel and hitting the second gadget from the
left on the Control Screen strip.  This is the parking button.  From then
on this is where the Control Screen returns to.  Whenever you want to
access the Control Screen, hit the Gadget showing two arrows at left edge
of the top strip of the Control Screen.  Up it comes.  Hit it again, and it
parks back down the bottom.

GOING FURTHER STILL - DIFFERENT TYPES

15.  Different item types.  Bring up the control screen.  Look at the KB
panel.  It has two listview gadgets.  The left-hand of these is a list of
item types (Free Bayesian, Free Integer, etc.).  Select 'Free Integer' by
hitting it with the LMB.  Send Control Screen away (hit the left-most two-
arrows button) to reveal the Easel.

16.  Scrollable Easel.  Ensure the Easel is active (step 10a).  Now move
the mouse over to the right hand side and keep moving.  The Easel scrolls
smoothly to the left - a tribute to the Amiga's hardware-assisted
scrolling.  This allows you to traverse a large knowledge base - and, since
scrolling happens even during drawing, allows you to draw very long links.
But, for now, we use it to gain empty space for a new inference net.

17a.  Draw a new box on the Easel in some empty space (= step 4) - it
should say 'Integer' in it and be a different colour from the Bayesian
boxes you had before.  Give it a name/label of 'Result' (= step 9).

17b.  Draw three antecedent boxes for it (to its left), and link them to
the Result.  Give them three different names e.g. 'I1', 'I2', 'I3'.  (Note
that this little inference net is not connected to the earlier one you
drew.)

18.  Run this wee inference net.  (Step 12: click over Result box with LMB
to bring up its Action panel and hit Reset then Infer.)  The User Question
panels will now show, not a slider value gadget, but a rectangle which says
'0' and into which you can type a number.  In each of the three that comes
up type a number, say 5, 6 and 7.

19.  Hit the Data button on Result's Action panel, and its Attribute
Details panel should appear, showing as its value the sum of the numbers
you entered (18 for the above numbers).

20.  To run it again there is a short cut to the Action panel, the OK-Act
button near its bottom.  Hit that.  (It short-cuts the process of hitting
OK and then finding the box and clicking on it with LMB.)  The Attribute
Action and Attribute Details panels form a pair, and each is accessible
from the other: the OK-Act button brings up the Action panel and the Data
button thereon brings up the Details panel.  Run it again, making sure this
time that the three numbers you enter are all above 1.

21.  Different Inference Method.  When the question sequence has stopped,
hit the Data button to bring up the Details panel.  On about the fourth row
you should see a radio button saying 'Inferred' and to its right a string
in a slot that says "X = A + B, C, ...".  This is the inference method by
which the values of all the antecedents are combined to give this
attribute's value.  "X = A + B, C, ..." means "Add all the antecedent
values together - X refers to this current attribute (labelled 'Result'), A
refers to the first antecedent, B to the second, and so on.  You can select
another inference method: attached at the left end of the slot is a wee
button; hit it with LMB.  Up comes a list of inference methods.  Select
multiplication, shown by "X = A * B, C, ...".  Hit the OK button on this
list to send it away, and the string in your slot should be updated
accordingly.

22.  Now run the inference net again.  (Step 20: notice how convenient the
OK-Act button is when you are changing things like inference method and
trying it out.)  Make sure all numbers are above 1.  Look at the Result
Data.  (If you entered 5, 6, 7 the result value should now be 210.)

23.  Suppressing irrelevant information.  Note that Istar only asks
questions that it needs to, and thus avoids asking irrelevant ones.  With
multiplication this happens if one of the antecedents is zero - we then
know the consequent will be zero, whatever values the remaining antecedents
take, so there is no point asking them.  Try it: run the inference net
again (step 20) and enter the second number as 0.  It should not then ask
the third number.

23a.  (There are various item types for which this effect can occur.
Integers with multiply, Booleans with AND or OR, and anything with Chooser
inference method, etc.)

24.  Short-cut to changing item type.  The main way of selecting item type
for your next box is to bring up the control screen, hit the item type
listview gadget for the type you want.  A bit of a nuisance; there is a
short-cut if you have a box of that type already showing on the Easel:
press 'T' for Type over that box and Istar will select its type.  You have
two types, Bayesian and Integer, with (Free) Integer currently selected.
Ensure the Easel is active (step 10a).  Then move the mouse pointer over
one of your earlier Bayesian boxes and press the T key.  The item type
should select Free Bayesian.  If you now draw a box it will be of Bayesian
type.

25.  Automatic Type Conversion.  Find some spare space on the Easel (step
16), create a (Free) Probability box to the left and an Integer to the
right.  Link the Bayesian box as antecedent of the Integer.  Run this
little inference net; you should be presented with a slider in the User
Question panel.  Slide the Probability to some suitable number, say 90.
Look at the value in the Integer attribute; it should say 90 (though early
versions of Istar might be one out due to rounding errors!).  As far as
possible Istar intelligently converts from one value type to another.

26.  Some errors.  Don't be scared of errors.  Try experiencing a few
errors that might happen - if you have not already done so!  Most do not do
any harm.

      #  Separating mouse from link end.  You might have noticed that
      occasionally after starting to draw link the mouse pointer is
      separated slightly from the end of the link.  It happens because
      there is a very slight delay between pressing of LMB and appearance
      of the link, and in this delay the mouse could have moved away
      slightly.  Note that it is the end of the link that is the important,
      active point, not the mouse pointer.  So it is the end of the link
      that should be dropped into the target box, rather than the mouse
      pointer.  Yhis feature is deliberate as it allows the mouse pointer
      to be moved out of the way of the active point, so you can see more
      clearly exactly where the end of the link is going.  Useful in a
      cluttered Easel.  (The same can also happen with drawing boxes,
      bending links, or whatever.)

      #  Link not formed.  In drawing a link, if you release LMB such that
      the end of the link is over empty space then it will not be formed -
      because it must link to something.  This can happen when link end and
      pointer are separated and you place the pointer rather than link end
      over the box.

      #  Cannot link item to itself.  With any box, start drawing a link
      from its right or left edge but release LMB while the end of the link
      is inside the same box.  An error panel appears.  Hit OK.

      #  Cannot create a cycle.  With one of the inference nets you have
      drawn, start drawing a link from the right-hand edge of the right-
      most box and (step 6b) bend it round to the left until it is over one
      of the antecedents, thus creating a cycle in the inference net.  This
      is not allowed; an error panel appears.  Hit OK.

      #  Invisible parts of net.  Some keys hide parts of the net.  e.g.
      the A or C keys show antecedents and consequents of a box.  Useful in
      simplifying a large net.  But if hit inadvertently can be confusing
      as you will see links ending in empty space etc.  If this has
      happened then hit the Enter Key on numeric keypad.  This restores
      all.


WHAT MORE IS THERE?

Lots more.  That should more than get you started.  But there's lots more:

      #  Larger inference nets
            #  Multi-step
            #  Loop prevention
            #  Scrolling
            #  Zooming
            #  Showing parts
      #  Goal lists
      #  Topics
      #  Find
      #  Forms
      #  Creating new item types, multiple attributes
      #  Creating new attribute types
      #  Document creation
      #  KB Dump, save, load.
      #  etc.


**  IF YOU WISH, YOU CAN NOW GO STRAIGHT TO <USING ISTAR>.


STARTING FROM CLI

To start Istar from CLI, simply type "Istar", maybe appending the version
number if there is one.  (I assume that you are familiar with paths etc. or
how to activate a program in another directory.)

Advanced use:
      From CLI you can start with various tracing on, by adding the
parameters:

      "-tm" to trace some activity in the modules mechanism
      "-tu" to trace some UA module activity
      "-te" to trace some activity among easels

so the command might be:

      Istar1.04 -tu -te

This will usually only be of value when you suspect some bug and want to
send me a trace of what happens.


STARTING WITH LOW MEMORY

The standard startup will just about start in a base 2 Mb A1200 as long as
your Workbench screen is simple PAL 640 by 256 and only 4 colours.  But in
other situations, from Version 1.04 onwards, you can take action to reduce
the amount of memory needed for your KB, before you create it.  There are
two ways of doing this.

LOW MEMORY METHOD 1:

This method involves starting up and immediately altering some of the
preferences via the Prefs panel.

      #  As soon as the Initial Control Panel comes up, click on the Prefs
      button to bring up the Preferences panel.

      #  Alter the MinDSA figure (usually 50000 longwords) to e.g. 10000;
      this reduces the size of a new KB from around 20000 bytes to around
      40000.

      #  Alter the size of the Easel from 1920 by 768 to something smaller,
      such as 640 by 256.  If its Depth is shown as 4 or higher reduce it
      to 3.  Doing these reduces the amount of memory for the Easel of a
      new KB.  The original size easel consumes around 200 kbytes for each
      bitplane, whereas 640 by 256 consumes only around 20 kbytes per
      bitplane.

Note, though, that you should only do this if you can work with a small KB
or easel.  Also, doing this will not alter the sizes of KB or easel that
are loaded from file.

LOW MEMORY METHOD 2:

Altering the preferences each time you load up can be tedious.  So Istar
can also load preferences from a file as it enters, which in version 1.04
is the file IstarSys:IstarPrefs (that is, IstarPrefs in the drawer to which
IstarSys: is assigned).  So all you need to do is supply such a file,
according to your needs.

      A file, IstarPrefs_LowMem is already supplied, so you need only enter
the following command from a CLI:

      COPY IstarSys:IstarPrefs_LowMem IstarSys:IstarPrefs

Bring up the Prefs panel to see the sizes it has given you.  If you wish
anything different then make up your own IstarPrefs file.  There are many
other Preferences that can be set via this file.  See 'Preferences' section
in the 'KBs' chapter for details.


WHAT IS AN ITEM?

An item in Istar is some focus of attention in the knowledge base.  It is
represented on the Easel by a box or group of boxes that move together.  An
item can have:

      #  any number of attributes (so far they have had only 1)
      #  relationships to other items

On the Item Type list on the KB panel, select 'Contract Meaning'.  This is
an item expressed by two boxes, when you draw it on the Easel.  The top one
represents the item itself, the bottom one an attribute of the item.

      You can define your own item types.  Let's try doing so.  Hit the
'New' button below the Item Type list.  A new panel, Create New Item Type
(CNIT) comes up for an item type called 'Unnamed'.  Change the name to
'Employee'.

      Now we will add attributes to the item type.  To add each attribute
to an item type we do three things:

      1.  Set the type
      2.  Set the name
      3.  Click the addition button just above the Type.

So first, hit the Type wee button, and select String from the list that
comes up; hit its OK button.  Then in the Attribute name to the left of the
Type, enter 'Name'.  Now hit the button above Type.  'Name' will appear in
the list in the middle of the panel.  Now create an attribute called 'Age',
of type Integer using the same steps.  Now create an attribute called 'Day
off' of Type 'Weekdays'.  Lastly, create an attribute called 'Salary' of
Type Integer.

      Then press OK (called Done in some versions) on the CNIT panel, and
it disappears.  Look at the SIT panel and you will see 'Employee' has been
added at the bottom and should be selected (if not, select it).

      Now draw on the Easel and you should get a tall box group with five
boxes: a smaller one at the top representing the employee itself, with four
attribute boxes: Name, Age, Day off, Salary.

      Notice the similarity to databases: item = entity or record,
attribute = attribute or field.  Unlike many databases new item types can
be created at any time.  And (probably not this version) new attributes can
be added to existing item types.


WHAT IS AN ATTRIBUTE?

An attribute is a holder of a value - such as Name, Age, etc.  It can be of
two types: in an item or a free attribute.  Each attribute is expressed by
a box.  When you started above, you were drawing (creating) free
attributes.  When you drew the Employee box group you created an item with
for attributes.  Free attributes actually belong to a hidden item, for
which there is no visible box.

      Attributes can be of a variety of types, which normally are defined
in terms of value types (see below).

      It became clear during the INCA research project that the
conventional picture, of items with several attributes, was not always
appropriate.  In an inference net a lot of items exist merely to support a
single attribute, and so Istar developed the idea of 'free attributes', in
which the item is still there for structural purposes but is de-emphasized
and not shown on the easel.  So in the Select Item Type panel the first set
of items is normally a load of free attributes.  They are represented by
single boxes.  Whether they are called attributes or items is immaterial.


Copyright (c) Andrew Basden, 1996