Re: Experimental aggregate_by/4 was dismissed (Was: India & France had their AI Bikini Moment)

I was exploring group_by/4 respectively a new
aggregate_by/4 for some machine learning statistics.
But the slowdown is not that aggravated if the extra
parameter _H is ground. The cost is smaller
 factor then. So I had a change of mind in favor
of more declarative aggregate/3. Too much new predicates
isn't healthy so I dismissed the idea of supporting
distinct/2, group_by/4 and a new aggregate_by/4.
 But somehow I fell in love with the idea of a new
firstof/2 predicate, instead of distinct/2, it could
be bootstrapped as follows:
```
firstof(X, Q) :-
    bagof(X, Q, L),
    L = [X|_].
```
Except it can be implemented like distinct/2 more
eagerly fitting to some other predicates from
the library(sequence):
```
p(1,a).
p(1,b).
p(2,c).
p(2,d).
p(2,e).
 ?- firstof(Y,p(X,Y)).
Y = a, X = 1;
Y = c, X = 2;
fail.
```
 Mild Shock schrieb:

Just noticed that group_by/4 calculates variables
and then delegates to bagof/3. But the later predicate
calculates also varables, so I suspect quite an overhead:
>
/* SWI-Prolog 9.3.19 */
group_by(By, Template, Goal, Bag) :-
     ordered_term_variables(Goal, GVars),
     ordered_term_variables(By+Template, UVars),
     ord_subtract(GVars, UVars, ExVars),
     bagof(Template, ExVars^Goal, Bag).
>
I went with another soluton. First I provided a variant
of aggregate/3 by the name aggregate_by/4 where one can
offload the internal term_variables/2 calculation.
Then use this bootstrapping:
>
/* Dogelog Player 1.3.0 */
group_by(Witness, Template, Goal, List) :-
    aggregate_by(Witness, bag(Template), Goal, List).
>
Here is some testing:
>
/* SWI-Prolog 9.3.19 */
?- length(_H,4000), time((between(1,2000,_),
         group_by(X,Y,(nonvar(_H),between(1,10,Y),between(1,10,X)),L),
         fail; true)).
% 1,153,998 inferences, 0.562 CPU in 0.568 seconds (99% CPU, 2051552 Lips)
true.
?- length(_H,8000), time((between(1,2000,_),
         group_by(X,Y,(nonvar(_H),between(1,10,Y),between(1,10,X)),L),
         fail; true)).
% 1,153,998 inferences, 1.047 CPU in 1.060 seconds (99% CPU, 1102326 Lips)
true.
>
/* Dogelog Player 1.3.0 */
?- length(_H,4000), time((between(1,2000,_),
         group_by(X,Y,(nonvar(_H),between(1,10,Y),between(1,10,X)),L),
         fail; true)).
% Zeit 399 ms, GC 0 ms, Lips 16987636, Uhr 10.02.2025 10:49
true.
?- length(_H,8000), time((between(1,2000,_),
         group_by(X,Y,(nonvar(_H),between(1,10,Y),between(1,10,X)),L),
         fail; true)).
% Zeit 400 ms, GC 1 ms, Lips 16945167, Uhr 10.02.2025 10:50
true.
>
The old version suffers from some term_variables/2
dependency whereas the new version is totally immune
on the size of the given goal, since any internal
term_variables/2 has been offloaded.
>
I couldn’t name aggregate_by/4 as aggregate/4, since
the later already exists in SWI-Prolog and SICStus Prolog
and has a different semantics, it is not the analog of
distinct/2, where one can specify Witnesses.
>
Mild Shock schrieb:

Hi,
>
India & France had their AI Bikini Moment.
Facinating behavior:
>
Macron Says He And PM Modi Will Push
https://www.youtube.com/watch?v=LwCK8yAnlkA
>
But don't be fooled, things are possibly
more connected:
>
Synthesia: France's 109-billion-euro AI investment
https://www.youtube.com/watch?v=_uyo4RG0Q6I
>
Bye
>
>
Mild Shock schrieb:

Hi,
>
Suddently I got an allergy to name a predicate
distinct/2. It is not so obvious that distinct/1 and
distinct/2 are related. There is no constant C such that:
>
distinct(X) :- distinct(C, X).
>
Just joking, but for some consistency with the introduction
of group_by/4 and aggregate_by/4 I went for the
name first_by/2. The name is more intuitive:
>
?- [user].
p(1,a).
p(1,b).
p(2,c).
p(2,d).
p(2,e).
^Z
true.
>
Now some queries:
>
?- p(X,Y), write(X-Y), nl, fail; true.
1-a
1-b
2-c
2-d
2-e
true.
>
?- first_by(X, p(X,Y)), write(X-Y), nl, fail; true.
1-a
2-c
true.
>
Cool! The name is also used here with the same semantics:
>
https://deephaven.io/core/docs/reference/table-operations/group-and- aggregate/firstBy/
>

>

>