Re: [opencog-dev] Is automatic alpha-conversion evil?

['Nil Geisweiller' via opencog]

On 11/17/2016 09:09 AM, Ben Goertzel wrote:

I think our conclusion in 2015 was that processes like the chainer
should do alpha conversion on the fly... so we should supply a utility
function making it easy for anyone to invoke alpha conversion in a
particular set of atoms..


The URE does alpha-conversion whenever necessary as to avoid name 
capture, there's no problem here. The problem comes from the fact that 
the atomspace may silently perform alpha-conversion when a scope link is 
added, ultimately there are solutions to address any problem resulting 
from such implicit conversions, but I can't help it, there's a part of 
me that feels ambivalent about it...


Anyway, I'll open a github issue whenever I encounter the problem I'm 
trying to describe in real life situation.


Nil




On Nov 17, 2016 3:54 PM, "'Nil Geisweiller' via opencog"
> wrote:

The reason I didn't want to open a bug report just yet is because I
wanted to discuss whether we really want to check alpha-equivalence
whenever a new scope link is added to the atomspace (and
subsequently consider both equal and return the previous one to the
user).

Nil

On 11/17/2016 05:13 AM, Linas Vepstas wrote:

Look, if there is a bug, open a bug report. I'm nervous about
these kind
of blanket statements about what's best and what is not best.  The
atomspace is already complicated, adding more complexity to it
is not a
solution.

--linas

On Wed, Nov 16, 2016 at 8:27 PM, Ben Goertzel 
>> wrote:

Nil,

I just reviewed our dialogue on this from a year ago...

It seems what we provisionally concluded there was that the
chainer
should do alpha-conversion on the fly in the course of pattern
matching ... but that the Atomspace shouldn't do
alpha-conversion
"automatically" in any other sense [unless we want to add
some Reduct
type engine on the Atomspace, which could do
alpha-conversion along
with other normalizations, but that becomes a separate issue]

We also discussed a cog-new-var command that could be used
to minimize
the complexities of alpha-conversions... (via reducing the
incidence
of redundant variable names)

In this case, the alpha-conversion done by the chainer in
the course
of doing its business, would need to handle LocalQuoteLink
correctly...

The choice of the chainer to do alpha-conversion but not
(yet) more
general types of reduction, would be made because
alpha-conversion is
cheaper and easier and of such broad utility.   Later
versions of the
chainer might do more general reductions as part of their
ordinary
business, as well...

I may be missing something; a year ago when William and I
were talking
about this, my head was fully immersed in the problem, but
it's less
the case right now...

-- Ben




On Wed, Nov 16, 2016 at 8:34 PM, 'Nil Geisweiller' via opencog

>> wrote:
> I'm back to this issue.
>
> The notion of LocalQuote is indeed incompatible with
systematic
> alpha-conversion.
>
> Consider this pattern
>
> (Get
>(VariableList
>   (TypedVariable
>  (Variable "$X")
>  (Type "TypedVariableLink"))
>   (TypedVariable
>  (Variable "$P")
>  (Type "PredicateNode"))
>   (TypedVariable
>  (Variable "$Q")
>  (Type "PredicateNode"))
>(LocalQuote
>   (ImplicationScope
>  (Variable "$X")
>  (Variable "$P")
>  (Variable "$Q"
>
> This fetches ImplicationScope links.
>
> If the following
>
> (ImplicationScope
>(Variable "$X")
>(Variable "$P")
>(Variable "$Q"))
>
> happen to be alpha-equivalent to something with different
variable
names it
> will render the Bind link invalid.
>
> Indeed alpha-conversion shouldn't be triggered in that
case, the
right idea
   

Re: [opencog-dev] Is automatic alpha-conversion evil?

[Linas Vepstas]

On Thu, Nov 17, 2016 at 12:54 AM, 'Nil Geisweiller' via opencog <
opencog@googlegroups.com> wrote:

> The reason I didn't want to open a bug report just yet is because I wanted
> to discuss whether we really want to check alpha-equivalence whenever a new
> scope link is added to the atomspace (and subsequently consider both equal
> and return the previous one to the user).
>

At the moment, I don't see any reasonable alternative.

You're suggesting that something is missing or that there's a problem, but
I haven't yet inferred what it is.   I did notice a comment in Rule.h
talking about a work-around to ScopeLinks -- perhaps that is a good
starting point.  What is it that you want to do, and why is the scope link
preventing you from doing it?

--linas


> Nil
>
> On 11/17/2016 05:13 AM, Linas Vepstas wrote:
>
>> Look, if there is a bug, open a bug report. I'm nervous about these kind
>> of blanket statements about what's best and what is not best.  The
>> atomspace is already complicated, adding more complexity to it is not a
>> solution.
>>
>> --linas
>>
>> On Wed, Nov 16, 2016 at 8:27 PM, Ben Goertzel > > wrote:
>>
>> Nil,
>>
>> I just reviewed our dialogue on this from a year ago...
>>
>> It seems what we provisionally concluded there was that the chainer
>> should do alpha-conversion on the fly in the course of pattern
>> matching ... but that the Atomspace shouldn't do alpha-conversion
>> "automatically" in any other sense [unless we want to add some Reduct
>> type engine on the Atomspace, which could do alpha-conversion along
>> with other normalizations, but that becomes a separate issue]
>>
>> We also discussed a cog-new-var command that could be used to minimize
>> the complexities of alpha-conversions... (via reducing the incidence
>> of redundant variable names)
>>
>> In this case, the alpha-conversion done by the chainer in the course
>> of doing its business, would need to handle LocalQuoteLink
>> correctly...
>>
>> The choice of the chainer to do alpha-conversion but not (yet) more
>> general types of reduction, would be made because alpha-conversion is
>> cheaper and easier and of such broad utility.   Later versions of the
>> chainer might do more general reductions as part of their ordinary
>> business, as well...
>>
>> I may be missing something; a year ago when William and I were talking
>> about this, my head was fully immersed in the problem, but it's less
>> the case right now...
>>
>> -- Ben
>>
>>
>>
>>
>> On Wed, Nov 16, 2016 at 8:34 PM, 'Nil Geisweiller' via opencog
>> > wrote:
>> > I'm back to this issue.
>> >
>> > The notion of LocalQuote is indeed incompatible with systematic
>> > alpha-conversion.
>> >
>> > Consider this pattern
>> >
>> > (Get
>> >(VariableList
>> >   (TypedVariable
>> >  (Variable "$X")
>> >  (Type "TypedVariableLink"))
>> >   (TypedVariable
>> >  (Variable "$P")
>> >  (Type "PredicateNode"))
>> >   (TypedVariable
>> >  (Variable "$Q")
>> >  (Type "PredicateNode"))
>> >(LocalQuote
>> >   (ImplicationScope
>> >  (Variable "$X")
>> >  (Variable "$P")
>> >  (Variable "$Q"
>> >
>> > This fetches ImplicationScope links.
>> >
>> > If the following
>> >
>> > (ImplicationScope
>> >(Variable "$X")
>> >(Variable "$P")
>> >(Variable "$Q"))
>> >
>> > happen to be alpha-equivalent to something with different variable
>> names it
>> > will render the Bind link invalid.
>> >
>> > Indeed alpha-conversion shouldn't be triggered in that case, the
>> right idea
>> > is that the ImplicationScope, when quoted corresponds to a
>> DIFFERENT atom
>> > than the one not being quoted. Also of course if we decide to not
>> perform
>> > systematic alpha-conversion then this problem doesn't arise.
>> >
>> > I'm re-iterating my question. Do we really want automatic
>> alpha-conversion
>> > to begin with?
>> >
>> > If the answer is yes then I suppose we need a way to tell that the
>> quoted
>> > version is different than then unquoted version.
>> >
>> > Nil
>> >
>> >
>> > On 10/22/2016 03:34 AM, Ben Goertzel wrote:
>> >>
>> >> Nil,
>> >>
>> >> Just brainstorming here, but perhaps the command for adding an Atom
>> >> should have an option that the user can set, which determines
>> whether
>> >> the results would be alpha-converted or not
>> >>
>> >> The default would be to do the alpha-conversion (which would be
>> >> appropriate if the variable names are say randomly generated, and
>> thus
>> >> of no particular 

Re: [opencog-dev] Is automatic alpha-conversion evil?

['Nil Geisweiller' via opencog]

Oh, my query was idiotic, $P and $Q wouldn't be typed PredicateNode, I 
meant something like EvaluationLink as to let a change to have the 
variables appear in them.


Nil

On 11/16/2016 07:56 PM, Linas Vepstas wrote:

Ah, the exception was due to a bug in the original email.  After fixing
that, it the code works just fine, I don't see what the problem is.  Cut
n paste of the works-for-me version below.

--linas

(use-modules (opencog) (opencog query) (opencog exec))

(define g
(Get
   (VariableList
  (TypedVariable
 (Variable "$X")
 (Type "TypedVariableLink"))
  (TypedVariable
 (Variable "$P")
 (Type "PredicateNode"))
  (TypedVariable
 (Variable "$Q")
 (Type "PredicateNode")))
   (LocalQuote
  (ImplicationScope
 (Variable "$X")
 (Variable "$P")
 (Variable "$Q")

;; except to find this
(ImplicationScope
   (TypedVariableLink
   (VariableNode "vvv")
   (TypeNode "ConceptNode")
   )
   (PredicateNode "ppp")
   (PredicateNode "qqq"))


(cog-execute! g)

;; returns the expected result. What is the problem?



On Wed, Nov 16, 2016 at 11:49 AM, Linas Vepstas > wrote:



On Wed, Nov 16, 2016 at 5:34 AM, 'Nil Geisweiller' via opencog
> wrote:

I'm back to this issue.

The notion of LocalQuote is indeed incompatible with systematic
alpha-conversion.

Consider this pattern

(Get
   (VariableList
  (TypedVariable
 (Variable "$X")
 (Type "TypedVariableLink"))
  (TypedVariable
 (Variable "$P")
 (Type "PredicateNode"))
  (TypedVariable
 (Variable "$Q")
 (Type "PredicateNode"))
   (LocalQuote
  (ImplicationScope
 (Variable "$X")
 (Variable "$P")
 (Variable "$Q"

This fetches ImplicationScope links.


Well, no, it throws an error:

 ERROR: Throw to key `C++-EXCEPTION' with args `("cog-new-link"
"Expected a VariableNode or a TypedVariableLink, got: LocalQuoteLink

(/home/linas/src/novamente/src/atomspace-git/opencog/atoms/core/VariableList.cc:61)")'.

That's with a pull from just now.

anyway, it is designed to fetch  ImplicationScopeLinks that are
ill-formed.  It's declared to find links like this:

(ImplicationScope
   (TypedVariableLink
   (TypeNode "ConceptNode")
   (VariableNode "vvv"))
   (PredicateNode "pa")
   (PredicateNode "qa"))

the variable never appears anywhere. Its -- well, the variable can
be completely discarded, and you'd get an equivalent
ImplicationScope that does not have any variables in it.


If the following

(ImplicationScope
   (Variable "$X")
   (Variable "$P")
   (Variable "$Q"))

happen to be alpha-equivalent to something with different
variable names it will render the Bind link invalid.


? I don't understand what you're saying. Why would it be "invalid"?
Is there a bug?  What do you mean by "happens to be
alpha-equivalent"? It is, by definition, alpha-equivalent to an
infinite number of other links.




Indeed alpha-conversion shouldn't be triggered in that case,


In which case? Alpha conversion of what?


the right idea is that the ImplicationScope, when quoted
corresponds to a DIFFERENT atom than the one not being quoted.


Can you clarify?


Also of course if we decide to not perform systematic
alpha-conversion then this problem doesn't arise.


You want to eliminate ScopeLink and Lambda Link ?  That's a very
serious change.



I'm re-iterating my question. Do we really want automatic
alpha-conversion to begin with?


What do you mean when you say "automatic"?  Either there is alpha
conversion, or there is not.   What else can it be?



If the answer is yes then I suppose we need a way to tell that
the quoted version is different than then unquoted version.


 what is different from what?  A better example is needed.

--linas



Nil


On 10/22/2016 03:34 AM, Ben Goertzel wrote:

Nil,

Just brainstorming here, but perhaps the command for adding
an Atom
should have an option that the user can set, which
determines whether
the results would be alpha-converted or not

The default would be to do the alpha-conversion (which would be
appropriate if the variable names are say randomly
generated, and thus
of no particular importance to the user -- the alpha
conversion is
then just preventing odd collisions between 

Re: [opencog-dev] Is automatic alpha-conversion evil?

[Ben Goertzel]

Nil,

I just reviewed our dialogue on this from a year ago...

It seems what we provisionally concluded there was that the chainer
should do alpha-conversion on the fly in the course of pattern
matching ... but that the Atomspace shouldn't do alpha-conversion
"automatically" in any other sense [unless we want to add some Reduct
type engine on the Atomspace, which could do alpha-conversion along
with other normalizations, but that becomes a separate issue]

We also discussed a cog-new-var command that could be used to minimize
the complexities of alpha-conversions... (via reducing the incidence
of redundant variable names)

In this case, the alpha-conversion done by the chainer in the course
of doing its business, would need to handle LocalQuoteLink
correctly...

The choice of the chainer to do alpha-conversion but not (yet) more
general types of reduction, would be made because alpha-conversion is
cheaper and easier and of such broad utility.   Later versions of the
chainer might do more general reductions as part of their ordinary
business, as well...

I may be missing something; a year ago when William and I were talking
about this, my head was fully immersed in the problem, but it's less
the case right now...

-- Ben




On Wed, Nov 16, 2016 at 8:34 PM, 'Nil Geisweiller' via opencog
 wrote:
> I'm back to this issue.
>
> The notion of LocalQuote is indeed incompatible with systematic
> alpha-conversion.
>
> Consider this pattern
>
> (Get
>(VariableList
>   (TypedVariable
>  (Variable "$X")
>  (Type "TypedVariableLink"))
>   (TypedVariable
>  (Variable "$P")
>  (Type "PredicateNode"))
>   (TypedVariable
>  (Variable "$Q")
>  (Type "PredicateNode"))
>(LocalQuote
>   (ImplicationScope
>  (Variable "$X")
>  (Variable "$P")
>  (Variable "$Q"
>
> This fetches ImplicationScope links.
>
> If the following
>
> (ImplicationScope
>(Variable "$X")
>(Variable "$P")
>(Variable "$Q"))
>
> happen to be alpha-equivalent to something with different variable names it
> will render the Bind link invalid.
>
> Indeed alpha-conversion shouldn't be triggered in that case, the right idea
> is that the ImplicationScope, when quoted corresponds to a DIFFERENT atom
> than the one not being quoted. Also of course if we decide to not perform
> systematic alpha-conversion then this problem doesn't arise.
>
> I'm re-iterating my question. Do we really want automatic alpha-conversion
> to begin with?
>
> If the answer is yes then I suppose we need a way to tell that the quoted
> version is different than then unquoted version.
>
> Nil
>
>
> On 10/22/2016 03:34 AM, Ben Goertzel wrote:
>>
>> Nil,
>>
>> Just brainstorming here, but perhaps the command for adding an Atom
>> should have an option that the user can set, which determines whether
>> the results would be alpha-converted or not
>>
>> The default would be to do the alpha-conversion (which would be
>> appropriate if the variable names are say randomly generated, and thus
>> of no particular importance to the user -- the alpha conversion is
>> then just preventing odd collisions between randomly generated
>> variable names created by two different processes)
>>
>> However, if the user wants they can override this default and specify
>> "no alpha conversion", and then it is their responsibility to check
>> and be sure their chosen VariableNode names are not going to be used
>> in a way that creates some conflict ...
>>
>> This option would need to be added to Scheme, python, Haskell
>> bindings, but also to the core API for adding scoped links, I guess...
>>
>> I am only about 83.456% sure I understand the problem here...
>>
>> -- Ben
>>
>>
>>
>> On Fri, Oct 21, 2016 at 11:55 PM, 'Nil Geisweiller' via opencog
>>  wrote:
>>>
>>> Hi,
>>>
>>> I start to think that automatic alpha-conversion is evil.
>>>
>>> First let me recall what it does. Say you've added
>>>
>>> (Scope (VariableList (Variable "$X") (Variable "$Y"))
>>> (And (Variable "$X") (Variable "$Y")))
>>>
>>> and you subsequently add
>>>
>>> (Scope (And (Variable "$gold") (Variable "$silver")))
>>>
>>> then recalling the handle of that last addition, you'd get the first
>>> alpha-equivalent scope, which is
>>>
>>> (Scope (VariableList (Variable "$X") (Variable "$Y"))
>>> (And (Variable "$X") (Variable "$Y")))
>>>
>>> This is rather confusing to the user, but even worse the pattern matcher
>>> behaves differently with the former or the latter. If you use the former
>>> to
>>> match grounds containing variables "$X" and "$Y" it may not work due to
>>> the
>>> pattern matcher discarding self-matches. The latter would match UNLESS
>>> the
>>> former has been previously added, because the variables "$gold" and
>>> "$silver" would be silently replaced by "$X" and "$Y". This is horribly
>>> confusing to the user!
>>>
>>> Second, it seems rather arbitrary to 

Re: [opencog-dev] Is automatic alpha-conversion evil?

[Linas Vepstas]

Ah, the exception was due to a bug in the original email.  After fixing
that, it the code works just fine, I don't see what the problem is.  Cut n
paste of the works-for-me version below.

--linas

(use-modules (opencog) (opencog query) (opencog exec))

(define g
(Get
   (VariableList
  (TypedVariable
 (Variable "$X")
 (Type "TypedVariableLink"))
  (TypedVariable
 (Variable "$P")
 (Type "PredicateNode"))
  (TypedVariable
 (Variable "$Q")
 (Type "PredicateNode")))
   (LocalQuote
  (ImplicationScope
 (Variable "$X")
 (Variable "$P")
 (Variable "$Q")

;; except to find this
(ImplicationScope
   (TypedVariableLink
   (VariableNode "vvv")
   (TypeNode "ConceptNode")
   )
   (PredicateNode "ppp")
   (PredicateNode "qqq"))


(cog-execute! g)

;; returns the expected result. What is the problem?



On Wed, Nov 16, 2016 at 11:49 AM, Linas Vepstas 
wrote:

>
>
> On Wed, Nov 16, 2016 at 5:34 AM, 'Nil Geisweiller' via opencog <
> opencog@googlegroups.com> wrote:
>
>> I'm back to this issue.
>>
>> The notion of LocalQuote is indeed incompatible with systematic
>> alpha-conversion.
>>
>> Consider this pattern
>>
>> (Get
>>(VariableList
>>   (TypedVariable
>>  (Variable "$X")
>>  (Type "TypedVariableLink"))
>>   (TypedVariable
>>  (Variable "$P")
>>  (Type "PredicateNode"))
>>   (TypedVariable
>>  (Variable "$Q")
>>  (Type "PredicateNode"))
>>(LocalQuote
>>   (ImplicationScope
>>  (Variable "$X")
>>  (Variable "$P")
>>  (Variable "$Q"
>>
>> This fetches ImplicationScope links.
>
>
> Well, no, it throws an error:
>
>  ERROR: Throw to key `C++-EXCEPTION' with args `("cog-new-link" "Expected
> a VariableNode or a TypedVariableLink, got: LocalQuoteLink
> (/home/linas/src/novamente/src/atomspace-git/opencog/
> atoms/core/VariableList.cc:61)")'.
>
> That's with a pull from just now.
>
> anyway, it is designed to fetch  ImplicationScopeLinks that are
> ill-formed.  It's declared to find links like this:
>
> (ImplicationScope
>(TypedVariableLink
>(TypeNode "ConceptNode")
>(VariableNode "vvv"))
>(PredicateNode "pa")
>(PredicateNode "qa"))
>
> the variable never appears anywhere. Its -- well, the variable can be
> completely discarded, and you'd get an equivalent ImplicationScope that
> does not have any variables in it.
>
>
>> If the following
>>
>> (ImplicationScope
>>(Variable "$X")
>>(Variable "$P")
>>(Variable "$Q"))
>>
>> happen to be alpha-equivalent to something with different variable names
>> it will render the Bind link invalid.
>>
>
> ? I don't understand what you're saying. Why would it be "invalid"? Is
> there a bug?  What do you mean by "happens to be alpha-equivalent"? It is,
> by definition, alpha-equivalent to an infinite number of other links.
>
>
>
>>
>> Indeed alpha-conversion shouldn't be triggered in that case,
>
>
> In which case? Alpha conversion of what?
>
>
>> the right idea is that the ImplicationScope, when quoted corresponds to a
>> DIFFERENT atom than the one not being quoted.
>
>
> Can you clarify?
>
>
>> Also of course if we decide to not perform systematic alpha-conversion
>> then this problem doesn't arise.
>>
>
> You want to eliminate ScopeLink and Lambda Link ?  That's a very serious
> change.
>
>
>>
>> I'm re-iterating my question. Do we really want automatic
>> alpha-conversion to begin with?
>>
>
> What do you mean when you say "automatic"?  Either there is alpha
> conversion, or there is not.   What else can it be?
>
>
>>
>> If the answer is yes then I suppose we need a way to tell that the quoted
>> version is different than then unquoted version.
>>
>
>  what is different from what?  A better example is needed.
>
> --linas
>
>
>>
>> Nil
>>
>>
>> On 10/22/2016 03:34 AM, Ben Goertzel wrote:
>>
>>> Nil,
>>>
>>> Just brainstorming here, but perhaps the command for adding an Atom
>>> should have an option that the user can set, which determines whether
>>> the results would be alpha-converted or not
>>>
>>> The default would be to do the alpha-conversion (which would be
>>> appropriate if the variable names are say randomly generated, and thus
>>> of no particular importance to the user -- the alpha conversion is
>>> then just preventing odd collisions between randomly generated
>>> variable names created by two different processes)
>>>
>>> However, if the user wants they can override this default and specify
>>> "no alpha conversion", and then it is their responsibility to check
>>> and be sure their chosen VariableNode names are not going to be used
>>> in a way that creates some conflict ...
>>>
>>> This option would need to be added to Scheme, python, Haskell
>>> bindings, but also to the core API for adding scoped links, I guess...
>>>
>>> I am only about 83.456% sure I understand the problem here...
>>>
>>> -- Ben
>>>
>>>
>>>

Re: [opencog-dev] Is automatic alpha-conversion evil?

['Nil Geisweiller' via opencog]

Another option would be to defined quoted atom types corresponding to 
all scope links, like


QuotedImplicationScopeLink

Or

ImplicationScopeLocalQuoteLink

which would inherit a LocalQuoteLink.

It has the advantage of being compact, unambiguous and circumvent the 
alpha-conversion issue altogether.


What do you think?

Nil

On 11/16/2016 01:34 PM, Nil Geisweiller wrote:

I'm back to this issue.

The notion of LocalQuote is indeed incompatible with systematic
alpha-conversion.

Consider this pattern

(Get
(VariableList
   (TypedVariable
  (Variable "$X")
  (Type "TypedVariableLink"))
   (TypedVariable
  (Variable "$P")
  (Type "PredicateNode"))
   (TypedVariable
  (Variable "$Q")
  (Type "PredicateNode"))
(LocalQuote
   (ImplicationScope
  (Variable "$X")
  (Variable "$P")
  (Variable "$Q"

This fetches ImplicationScope links.

If the following

(ImplicationScope
(Variable "$X")
(Variable "$P")
(Variable "$Q"))

happen to be alpha-equivalent to something with different variable names
it will render the Bind link invalid.

Indeed alpha-conversion shouldn't be triggered in that case, the right
idea is that the ImplicationScope, when quoted corresponds to a
DIFFERENT atom than the one not being quoted. Also of course if we
decide to not perform systematic alpha-conversion then this problem
doesn't arise.

I'm re-iterating my question. Do we really want automatic
alpha-conversion to begin with?

If the answer is yes then I suppose we need a way to tell that the
quoted version is different than then unquoted version.

Nil

On 10/22/2016 03:34 AM, Ben Goertzel wrote:

Nil,

Just brainstorming here, but perhaps the command for adding an Atom
should have an option that the user can set, which determines whether
the results would be alpha-converted or not

The default would be to do the alpha-conversion (which would be
appropriate if the variable names are say randomly generated, and thus
of no particular importance to the user -- the alpha conversion is
then just preventing odd collisions between randomly generated
variable names created by two different processes)

However, if the user wants they can override this default and specify
"no alpha conversion", and then it is their responsibility to check
and be sure their chosen VariableNode names are not going to be used
in a way that creates some conflict ...

This option would need to be added to Scheme, python, Haskell
bindings, but also to the core API for adding scoped links, I guess...

I am only about 83.456% sure I understand the problem here...

-- Ben



On Fri, Oct 21, 2016 at 11:55 PM, 'Nil Geisweiller' via opencog
 wrote:

Hi,

I start to think that automatic alpha-conversion is evil.

First let me recall what it does. Say you've added

(Scope (VariableList (Variable "$X") (Variable "$Y"))
(And (Variable "$X") (Variable "$Y")))

and you subsequently add

(Scope (And (Variable "$gold") (Variable "$silver")))

then recalling the handle of that last addition, you'd get the first
alpha-equivalent scope, which is

(Scope (VariableList (Variable "$X") (Variable "$Y"))
(And (Variable "$X") (Variable "$Y")))

This is rather confusing to the user, but even worse the pattern matcher
behaves differently with the former or the latter. If you use the
former to
match grounds containing variables "$X" and "$Y" it may not work due
to the
pattern matcher discarding self-matches. The latter would match
UNLESS the
former has been previously added, because the variables "$gold" and
"$silver" would be silently replaced by "$X" and "$Y". This is horribly
confusing to the user!

Second, it seems rather arbitrary to try to detect this kind of
equivalence
while there's an infinity of others. For instance

(And (Variable "$X") (And (Variable "$Y"))

is equivalent to

(And (Variable "$X") (Variable "$Y"))

For these reasons I think semantic equivalence detection shouldn't be
incorporated into the AtomSpace. The AtomSpace should take care of the
syntax only (OK, with the exception of unordered links), as it's always
been, and this task should differed to another process working above the
AtomSpace.

It was suggested a while ago to have a normal form reduction engine
for the
AtomSpace, similar to MOSES', and such an engine could be used to reduce
while adding atoms, if the user chooses so. This is a much cleaner
way to
handle that. Also since semantic equivalence is undecidable, there will
always be a battle between completeness and performance. Another
reason to
have this ever growing monster above the AtomSpace rather than in it.

OK, I don't know if I've convinced you, or even if I've convinced
myself,
but it's really a discussion we need to have.

Opinions welcome.

Nil

--
You received this message because you are subscribed to the Google
Groups
"opencog" group.
To unsubscribe from this group and stop receiving emails 

Re: [opencog-dev] Is automatic alpha-conversion evil?

['Nil Geisweiller' via opencog]

I'm back to this issue.

The notion of LocalQuote is indeed incompatible with systematic 
alpha-conversion.


Consider this pattern

(Get
   (VariableList
  (TypedVariable
 (Variable "$X")
 (Type "TypedVariableLink"))
  (TypedVariable
 (Variable "$P")
 (Type "PredicateNode"))
  (TypedVariable
 (Variable "$Q")
 (Type "PredicateNode"))
   (LocalQuote
  (ImplicationScope
 (Variable "$X")
 (Variable "$P")
 (Variable "$Q"

This fetches ImplicationScope links.

If the following

(ImplicationScope
   (Variable "$X")
   (Variable "$P")
   (Variable "$Q"))

happen to be alpha-equivalent to something with different variable names 
it will render the Bind link invalid.


Indeed alpha-conversion shouldn't be triggered in that case, the right 
idea is that the ImplicationScope, when quoted corresponds to a 
DIFFERENT atom than the one not being quoted. Also of course if we 
decide to not perform systematic alpha-conversion then this problem 
doesn't arise.


I'm re-iterating my question. Do we really want automatic 
alpha-conversion to begin with?


If the answer is yes then I suppose we need a way to tell that the 
quoted version is different than then unquoted version.


Nil

On 10/22/2016 03:34 AM, Ben Goertzel wrote:

Nil,

Just brainstorming here, but perhaps the command for adding an Atom
should have an option that the user can set, which determines whether
the results would be alpha-converted or not

The default would be to do the alpha-conversion (which would be
appropriate if the variable names are say randomly generated, and thus
of no particular importance to the user -- the alpha conversion is
then just preventing odd collisions between randomly generated
variable names created by two different processes)

However, if the user wants they can override this default and specify
"no alpha conversion", and then it is their responsibility to check
and be sure their chosen VariableNode names are not going to be used
in a way that creates some conflict ...

This option would need to be added to Scheme, python, Haskell
bindings, but also to the core API for adding scoped links, I guess...

I am only about 83.456% sure I understand the problem here...

-- Ben



On Fri, Oct 21, 2016 at 11:55 PM, 'Nil Geisweiller' via opencog
 wrote:

Hi,

I start to think that automatic alpha-conversion is evil.

First let me recall what it does. Say you've added

(Scope (VariableList (Variable "$X") (Variable "$Y"))
(And (Variable "$X") (Variable "$Y")))

and you subsequently add

(Scope (And (Variable "$gold") (Variable "$silver")))

then recalling the handle of that last addition, you'd get the first
alpha-equivalent scope, which is

(Scope (VariableList (Variable "$X") (Variable "$Y"))
(And (Variable "$X") (Variable "$Y")))

This is rather confusing to the user, but even worse the pattern matcher
behaves differently with the former or the latter. If you use the former to
match grounds containing variables "$X" and "$Y" it may not work due to the
pattern matcher discarding self-matches. The latter would match UNLESS the
former has been previously added, because the variables "$gold" and
"$silver" would be silently replaced by "$X" and "$Y". This is horribly
confusing to the user!

Second, it seems rather arbitrary to try to detect this kind of equivalence
while there's an infinity of others. For instance

(And (Variable "$X") (And (Variable "$Y"))

is equivalent to

(And (Variable "$X") (Variable "$Y"))

For these reasons I think semantic equivalence detection shouldn't be
incorporated into the AtomSpace. The AtomSpace should take care of the
syntax only (OK, with the exception of unordered links), as it's always
been, and this task should differed to another process working above the
AtomSpace.

It was suggested a while ago to have a normal form reduction engine for the
AtomSpace, similar to MOSES', and such an engine could be used to reduce
while adding atoms, if the user chooses so. This is a much cleaner way to
handle that. Also since semantic equivalence is undecidable, there will
always be a battle between completeness and performance. Another reason to
have this ever growing monster above the AtomSpace rather than in it.

OK, I don't know if I've convinced you, or even if I've convinced myself,
but it's really a discussion we need to have.

Opinions welcome.

Nil

--
You received this message because you are subscribed to the Google Groups
"opencog" group.
To unsubscribe from this group and stop receiving emails from it, send an
email to opencog+unsubscr...@googlegroups.com.
To post to this group, send email to opencog@googlegroups.com.
Visit this group at https://groups.google.com/group/opencog.
To view this discussion on the web visit
https://groups.google.com/d/msgid/opencog/580A3A75.1020708%40gmail.com.
For more options, visit https://groups.google.com/d/optout.






--
You received