Learning ordered/3

:-module(ordered, [background_knowledge/2
                  ,metarules/2
                  ,positive_example/2
                  ,negative_example/2
                  ]).


:-use_module(configuration).

/** <module> A longer example for Veedrac on HN.

Initiate a multi-predicate learning session to learn all the targets
jointly. Use dynamic learning to learn each target incrementally, by
reusing each clause learned in an earlier iteration:

==
?- time(learn_dynamic([llength/2,shorter/2,ordered/3])).
llength([],0).
llength(A,B):-tail(A,C),p(B,D),llength(C,D).
shorter(A,B):-llength(A,C),llength(B,D),s(C,D).
ordered(A,B,C):-shorter(A,B),shorter(B,C).
% 19,682 inferences, 0.016 CPU in 0.009 seconds (172% CPU, 1259648 Lips)
true.
==

If something fails, check that the following configuration options are
set:

==
?- list_config.
example_clauses(call)
experiment_file(data/drafts/functions/ordered.pl,ordered)
learner(louise)
max_invented(0)
minimal_program_size(2,inf)
recursion_depth_limit(dynamic_learning,50000)
recursive_reduction(false)
reduction(plotkins)
resolutions(5000)
symbol_range(predicate,[P,Q,R,S,T])
symbol_range(variable,[X,Y,Z,U,V,W])
theorem_prover(resolution)
unfold_invented(false)
true.
==

*/

% ================================================================================
% Metarule constraints. This is actually a misnomer- those are _clause_
% constraints, used to eliminate unwanted clauses. We use them here
% mainly for aesthetical reasons.
% ================================================================================

% This constraint cuts left-recursive shorter/2 clauses that make it
% fiddly to run the learned hypothesis in Prolog.
configuration:metarule_constraints(M,fail):-
    M =.. [m,Id,P|Ps]
    ,Id \= projection
    ,left_recursive(P,Ps).

left_recursive(T,[T|_Ps]):-
    !.
left_recursive(T,[T,T|_Ps]):-
    !.

% Cuts clause re-definitng llength/2 by shorter/2:
% llength(A,B):-shorter(A,C),s(B,D),llength(C,D).
% This would be fine in an exploratory session, say.
configuration:metarule_constraints(M,fail):-
    M =.. [m,_Id,llength,shorter|_].


% Avoids definition of shorter/2 by tail/2:
% shorter(A,B):-tail(A,C),tail(B,D),shorter(C,D).
% This compares the tails of two lists, which is the same as comparing
% the full lists, so it's correct, but again it's more useful in an
% exploratory session.
configuration:metarule_constraints(M,fail):-
    M =.. [m,_Id,shorter,tail|_].


% ================================================================================
% Various configs
% ================================================================================

% Sets the maximum number of predicates allowed to be invented.
% It's set to 0 because we don't need to invent any. We still use
% dynamic learning for for the ability to learn incrementally by reusing
% predicates learned (or invented) in earlier learning steps.
:- auxiliaries:set_configuration_option(max_invented, [0]).


% ================================================================================
% Declarations of Meta-Interpretive Learning problem elements
% Elements of a MIL problem are: metarules, background knowledge (BK)
% and examples.
% ================================================================================

% Metarules for the experiment are derived from Prolog program
% skeletons. They are essentially "program recipes" that can be reused
% to construct different programs.
configuration: list_rec_func metarule 'P(x,y):- Q(x,z),R(y,u),P(z,u)'.
configuration: triadic_chain metarule 'P(x,z,y):- Q(x,z), R(z,y)'.
configuration: m1 metarule 'P(x,y):- Q(x,z), R(y,u), S(z,u)'.

% Metarule declarations - though specific metarules are declared for
% each learning target, multi-predicate learning allows Louise to use
% any of the declared metarules for any target.
metarules(llength/2,[abduce,list_rec_func]).
metarules(shorter/2,[m1]).
metarules(ordered/3,[triadic_chain]).

% Background knowledge (BK) declarations - again, multi-predicate
% learning means any BK predicate can be used for any target. For
% example, we could leave the two latter lists empty and add s/2 to the
% first.
background_knowledge(llength/2, [tail/2,p/2]).
background_knowledge(shorter/2, [s/2]).
background_knowledge(ordered/3, [s/2
                                ]).

% Positive examples are declared separately for each target. They are
% added to the BK during learning, so any learning target can be defined
% in terms of these partial, extensional definitions of other targets.
positive_example(llength/2,E):-
        member(E, [llength([],0)
                  ,llength([a],s(0))
                  %,llength([a,b],s(s(0)))
                  ]
              ).
positive_example(shorter/2,shorter(Xs,Ys)):-
        member(Xs-Ys,[[a]-[b,c]
             ,[1,2]-[3,4,5]
                     ]
              ).
positive_example(ordered/3,ordered(S1,S2,S3)):-
        member([S1,S2,S3],[[[a],[b,c],[d,e,f]]
                          ]
              ).

% Negative examples
negative_example(shorter/2,_):-
        fail.
negative_example(llength/2,_):-
        fail.
negative_example(ordered/3,ordered(S1,S2,S3)):-
% These are not strictly needed - but I add them to address Veedrac's
% concern that Louise is not learning a sound program for ordered/2.
        member([S1,S2,S3],[[[a],[b],[c]]
              ,[[a,b,c],[a,b],[c]]
              ]
              ).


% ================================================================================
% Background knowledge predicate definitions.
%
% BK in Louise consists of arbitrary Prolog programs. Strictly speaking,
% they must have definite datalog heads, but this is not enforced in
% practice. I like to see BK as a kind of programming library,
% containing sub-routines, sub-programs from which the target programs
% can be composed.
%
% Metarules on the other hand must be definite datalog and Louise _does_
% enforce this restriction. The predicates below are used to allow
% Louise to process non-datalog examples, i.e. examples with function
% symbols in their head or body literals. In a sense, the function
% symbols are "hidden away" in the BK. The technique is similar to
% "flattening": Rouveirol, C, Machine Learning, 1994:
%
% https://paperity.org/p/7456046/flattening-and-saturation-two-representation-changes-for-generalization
%
% The predicates below have universal applicability for problems
% involving lists and numbers. head/2 and tail/2 split a list, s/2 and
% p/2 are used to "dereference" Peano number functions.
%
% Note hat head/2 is not actually used above.
% ================================================================================

%!      head(?List,?Head) is det.
%
%       True when Head is the head of list List.
%
head([H|_T],H).

%!      tail(?List,?Tail) is det.
%
%       True when Tail is the tail of list List.
%
tail([_H|T],T).


%!      p(?Number, ?Predecessor) is nondet.
%
%       A Peano Number and its Predecessor.
%
p(s(N),N).


%!      s(?Number, ?Successor) is nondet.
%
%       A Peano Number and its Successor.
%
s(N,s(N)).