logic-and-learning-lab/Popper

Can Popper use constants?

Closed this issue · 6 comments

pudumagico commented

Simple question as the header says.
I would like to learn rules that have constants in them.

andrewcropper commented

No, the only way to learn rules with constants is to create a monadic relation for each possible constant symbol, like so:

constant(agent_white, agent).
constant(agent_black, agent).
constant(int_0, int).
constant(int_1, int).
constant(int_2, int).


body_pred(P,1):-
    constant(P,_).

type(P,(T,)):-
    constant(P,T).

This bias file shows a full example: https://github.com/logic-and-learning-lab/Popper/blob/c1935cfa98112debe91ef9295c3d260f229a22d0/examples/iggp-gt-chicken-goal/bias.pl

There is a version of Popper, called MagicPopper, that can learn rules with constants (including from infinite domains) but I have still not merged it into the main repository: https://github.com/celinehocquette/magicpopper?tab=readme-ov-file

pudumagico commented

Thank you! magicpopper worked right away :)
I will try to use popper as well.
The performance should be similar?

  • 👍1
andrewcropper commented
pudumagico commented

Thank you for the prompt reply.
I am having trouble running it on Popper.

############# bias ##############

head_pred(hard, 1).

body_pred(contains, 2).
body_pred(direction, 2).
body_pred(shape, 2).
body_pred(size, 2).

constant(small, szid).
constant(large, szid).

constant(airliner, shid).
constant(biplane, shid).
constant(jet, shid).
constant(fighter, shid).
constant(utility, shid).
constant(tandem, shid).
constant(road, shid).
constant(mountain, shid).
constant(articulated, shid).
constant(double, shid).
constant(regular, shid).
constant(school, shid).
constant(truck, shid).
constant(suv, shid).
constant(minivan, shid).
constant(sedan, shid).
constant(wagon, shid).
constant(chopper, shid).
constant(scooter, shid).
constant(cruiser, shid).
constant(dirtbike, shid).

constant(south, did).
constant(southwest, did).
constant(west, did).
constant(northeast, did).
constant(north, did).
constant(northwest, did).
constant(west, did).
constant(southeast, did).

type(hard, (sid,)).
type(contains, (sid, oid)).
type(direction, (oid, did)).
type(shape, (oid, shid)).
type(size, (oid, szid)).

body_pred(P,1):-
constant(P,_).

type(P,(T,)):-
constant(P,T).

####### bk ##############

:-style_check(-discontiguous).

contains(s1, o1_0).
direction(o1_0, southeast).
shape(o1_0, double).
size(o1_0, small).
...

examples

pos(hard(s1)).
pos(hard(s2)).
neg(hard(s4)).
neg(hard(s8)).
...

It throws the error:

hard(V0):- small(V2),size(V1,V2),contains(V0,V1).
Traceback (most recent call last):
File "/home/nelson/Research/ILP4SD/Popper/popper.py", line 8, in
prog, score, stats = learn_solution(settings)
File "/home/nelson/Research/ILP4SD/Popper/popper/loop.py", line 1729, in learn_solution
timeout(settings, popper, (settings, tester, bkcons), timeout_duration=int(settings.timeout-time_so_far),)
File "/home/nelson/Research/ILP4SD/Popper/popper/util.py", line 81, in timeout
result = func(*args, **kwargs)
File "/home/nelson/Research/ILP4SD/Popper/popper/loop.py", line 1658, in popper
Popper(settings, tester, bkcons).run()
File "/home/nelson/Research/ILP4SD/Popper/popper/loop.py", line 247, in run
pos_covered = tester.test_prog_pos(prog)
File "/home/nelson/Research/ILP4SD/Popper/popper/tester.py", line 113, in test_prog_pos
return frozenset(query_once(q)['S'])
File "/home/nelson/miniconda3/lib/python3.10/site-packages/janus_swi/janus.py", line 243, in query_once
return _swipl.call(query, inputs, keep)
janus.PrologError: ''/1: Unknown procedure: small/1

If possible could you help me spot the error?
Thank you again.

andrewcropper commented

For each constant, you need to define the corresponding relation in the BK file, such as:

airliner(airliner).
small(small).

It is silly, but I have not yet implemented the code to automatically add them to the BK (I will move it up my TODO list),

pudumagico commented

It worked!
Thank you so much for your help!