Thelma - a Theory Learning Machine for Meta-Interpretive Learning

Thelma is an Inductive Logic Programming system. In particular, it is an implementation of Meta-Interpretive Learning (Muggleton et al. 2014), similar to Metagol. It learns first order logic theories in the form of dyadic datalog definite programs. It is trained on examples given as ground unit clauses and with background knowledge given as arbitrary Prolog programs.

As a MIL implementation, Thelma can perform predicate invention and can learn recursive hypotheses, including hypotheses with mutually recursive clauses.

Example of use

Thelma runs on Swi-Prolog version 8.0.0 or later.

See the data/examples/ directory for examples.

Follow the steps below to run an example that learns the a^nb^n Context Free Grammar from three positive examples.

A first experiment: learning a grammar for the a^nb^n context-free language

1. Consult the project's load file to load the project:

   ?- [load].

   This should start the Swi-Prolog IDE and open the project source files in the editor, including configuration.pl used in the next step. It should also start the documentation server and open this README file in your browser.

2. Edit the configuration file:

   Edit configuration.pl in the Swi-Prolog editor (or your favourite text editor) and make sure the name of the current experiment file is set to anbn.pl:

   experiment_file('data/examples/anbn.pl',anbn).

   Also set appropriate upper limits for total clauses and invented clauses in the learned hypothesis:

   depth_limits(3,1).

3. Recompile the project:

   ?- make.

4. Run the following query to train Thelma on the data in anbn.pl and print the results to the console.

   ?- learn('S'/2).
   % Clauses: 1; Invented: 0
   % Clauses: 2; Invented: 0
   % Clauses: 2; Invented: 1
   % Clauses: 3; Invented: 0
   % Clauses: 3; Invented: 1
   'S'(A,B):-'A'(A,C),'B'(C,B).
   'S'(A,B):-'S_1'(A,C),'B'(C,B).
   'S_1'(A,B):-'A'(A,C),'S'(C,B).
   true .

Understanding the training results

The hypothesis learned in the a^nb^n experiment, above translates to the following context-free grammar, in Definite Clause Grammars notation (DCG):

'S' --> 'A', 'B'.
'S' --> 'S_1', 'B'.
'S_1' --> 'A', 'S'.

Or, in BNF notation:

<S> ::= <A> <B>
<S> ::= <S_1> <B>
<S_1> ::= <A> <S>

This is a general definition of the a^nb^n grammar that covers all strings in the language and no strings outside the language, for arbitrary n.

In the learned hypothesis the predicate 'S_1'/2 is invented. Although it is not given in the background knowledge, it is necessary to complete the learning process. It is reconstructed from existing background knowledge and metarules, during learning. Note also that 'S_1'/2 is mutually recursive with 'S'/2.

The grammar learned by Thelma is a Prolog program and so it can be run both as a recogniser and a generator. See Context Free Grammars for an explanation of how a DCG definition of a^nb^n like the one learned by Thelma works.

Run as a generator

At the end of training the learned hypothesis is only printed at the Swi-Prolog top-level. In order to use it you first have to save it in a file and then consult the file, to load its clauses into memory.

You can always copy/paste the learned hypothesis into anbn.pl itself. Remember to save and reload it afterwards.

Once this is done, you can run the learned hypothesis as a generator by leaving its first variable unbound. For example, if you saved the learned grammar to anbn.pl, run the following query:

?- anbn:'S'(A,[]).
A = [a, b] ;
A = [a, a, b, b] ;
A = [a, a, a, b, b, b] ;
A = [a, a, a, a, b, b, b, b] ;
A = [a, a, a, a, a, b, b, b, b|...] .

If you saved the learned grammar in a different file, instead replace the module qualifier "anbn:" with that file's module name followed by a colon. If the file is not a module, simply call 'S'(A, []) without a module qualifier.

Test correctness

To test the grammar correctly recognises a^nb^n strings up to n = 100,000, run this query (assuming you have saved the grammar in the module 'anbn.pl'):

?- _N = 100_000, findall(a, between(1,_N,_), _As), findall(b, between(1,_N,_),_Bs), append(_As,_Bs,_AsBs), anbn:'S'(_AsBs,[]).
true .

You can try higher numbers up to the limits of your computational resources. The grammar is correct for any n.

Does the learned theory recognise strings it shouldn't? Try these tests:

% Not empty.
?- anbn:'S'([],[]).
false.

% Not only a
?- anbn:'S'([a],[]).
false.

% Not only b
?- anbn:'S'([b],[]).
false.

% Not starting with b
?- anbn:'S'([b|_],[]).
false.

% Not more a's than b's.
?- anbn:'S'([a,a,b],[]).
false.

% Not more b's than a's.
?- anbn:'S'([a,b,b],[]).
false.

The a^nb^n experiment file.

The contents of the experiment file used in the first experiment above, anbn.pl, are as follows (excluding comments):

configuration:metarule(unchain, [P,Q,R], [X,Y,Z], (mec(P,X,Y) :- mec(Q,X,Z), mec(R,Z,Y))).
configuration:order_constraints(unchain,[P,Q,_R],[X,Y,Z],[P>Q],[X>Z,Z>Y]).

background_knowledge('S'/2,['A'/2,'B'/2]).

metarules('S'/2,[unchain]).

positive_example('S'/2,E):-
	member(E, ['S'([a,b],[])
		  ,'S'([a,a,b,b],[])
		  ,'S'([a,a,a,b,b,b],[])
		  ]).

negative_example('S'/2,_):-
	fail.

'A'([a|A], A).
'B'([b|A], A).

Metarules and order constraints

The first two lines declare a new metarule called unchain and its order constraints. unchain is a version of the commonly used chain metarule, defined in configuration.pl, having one less lexicographic order constraint.

Metarules and order constraints are central to Meta-Interpretive Learning. The following is a high-level discussion in the context of their implementation in Thelma. For an in-depth discussion see the links in the bibliography section.

Metarules

Metarules are second-order definite Datalog clauses used as "templates" for clauses added to a hypothesis. In the metarule/4 clause in the anbn.pl experiment file, above, mec(P,X,Y) :- mec(Q,X,Z), mec(R,Z,Y) is an encapsulation of the metarule P(X,Y):- Q(X,Z), R(Z,Y) in a first order term ("mec" stands for "metarule encapsulation"). This encapsulation is necessary to allow second-order metarules to be processed by a first-order language like Prolog.

In the metarule/4 clause above, [P,Q,R] is the set of second order, existentially qualified variables and [X,Y,Z] the set of first order, universally quantified variables in unchain.

The existentially quantified variables in a metarule range over the set of predicate symbols including a) the predicate symbol of the target predicate, b) each of the predicate symbols of predicates in the background knowledge and c) any symbols invented during learning. In the a^nb^n experiment, the predicate symbol of the target predicate is 'S', the predicate symbols of the background knowledge are 'A' and 'B' and 'S_1' is the predicate symbol of an invented predicate.

The universally quantified variables in a metarule range over the constants in the examples of the target predicate and the bakcground knowledge (the Herbrand universe). In the a^nb^n experiment, the Herbrand universe is the set [a,b].

During the search for a hypothesis, the variables in a metarule are bound to predicate symbols and a metasubstitution is created. A metasubstitution is a substitution of existentially quantified variables for predicate symbols. For example, a metasubstitution created from the unchain metarule in the a^nb^n experiment might be:

{P/'S',Q/'A',R/'B'}

Where a term V/T means that the variable V is bound to the term T.

During learning, the body literals of a metarule instantiated by the metasubstitution are proved. If the proof succeeds, the metasubstitution is added to an abduction store. At the end of learning the metasubstitutions in the abduction store are projected onto their corresponding metarules to form the clause of the hypothesis.

For instance, the example metasubstitution above would be projected onto the unchain metarule to add the following clause to the learned hypothesis:

'S'(A,B):-'A'(A,C),'B'(C,B).

Order constraints

Order constraints guarantee termination and so allow learning of recursive theories. In the order_constraints/5 clause for unchain, in the anbn.pl experiment file, above, there are two sets of constraints: [P>Q] and [X>Z,Z>Y]. [P>Q] is a lexicographic order constraint and [X>Z,Z>Y] an interval inclusion constraint.

Ordering constraints require a total ordering of the predicate symbols and a total ordering of the constants in the Herbrand base of the predicates defined in the background knowledge. In Thelma, these orders are derived automatically from the order in which background predicates, their clauses and their terms, are encountered in an experiment file during compilation. Predicate symbols in the background knowledge are assigned a lexicographic order while their constants are assigned an interval inclusion order. See the structured documentation of the predicate order_constraints/3 in module src/auxiliaries.pl for an explanation of how this is done.

During the search for a hypothesis, termination is guaranteed because atoms with a higher order are proved by the meta-interpreter before atoms with a lower order. The two orders are finite and so cannot descend infinitely. See (Knuth and Bendix, 1970) for the original description and (Zhang et al. 2005) for a more recent discussion.

In unchain, the lexicographic constraint [P>Q] means that any predicate symbol bound to the variable P in unchain must be above any predicate symbol bound to the variable 'Q', with respect to the lexicographic ordering over the Herbrand base.

Accordingly, the intervalinclusion constraint [X>Z,Z>Y] means that any constant bound to the variable X in unchain must be above any constant bound to the variable X and so on for Z>Y, with respect to the interval inclusion order over the Herbrand base.

You can query order_constraints/3 to inspect the total orders it derives from an experiment file. For example, for the anbn.pl experiment file, the following orders are derived:

?- order_constraints('S'/2,Ps,Cs).
Ps = ['A'/2, 'B'/2],
Cs = [[a|_9572], [b|_9560], []]

Above, the two lists, Ps and Cs represent the lexicographic and interval orders, respectively. In this representation, a term p is above another term q in their respective order, iff p is before q in an ordering list.

The symbol of the target predicate is always added to the start of the lexicographic order list (this is done after order_constraints/2 is called, so you cannot see it in the query above). Additionally, any predicates during learning are added to the lexicographic order list right after the target predicate. The complete lexicographic order list for the anbn.pl experiment file with the invented predicate symbol S_1/2 would look like the following:

['S'/2, 'S_1'/2, 'A'/2, 'B'/2]

Given the two orders above and the lexicographic constraint P>Q in unchain, the metasubstitution {P/'S',Q/'A',R/'B'} upholds the constraint, whereas a metasubstitution {P/B,Q/'A',R/'S'} violates the constraint because it binds P to 'B' and Q to 'A' when P>Q and A/2>B/2.

Violation of a constraint means that the corresponding metasubstitution is not allowed and a new metasubstitution is tried. This way, the search for hypotheses only considers hypotheses with clauses that satisfy the constraints and are guaranteed to terminate.

Background knowledge and examples

ILP algorithms are characterised by their use of background knowledge to constraint the search for hypotheses.

The third line of code in anbn.pl declares the background knowledge for the target predicate 'S'/2. 'S'/2 is the start symbol of the grammar. 'A'/2 and 'B'/2, the terminals in the a^nb^n language, are given as background knowledge in the experiment file.

Conceptually, the function of the background knowledge is to constrain the search to hypotheses that satisfy the cirterion that the background knowledge definitions and the hypothesis must entail all positive examples, and none of the negative examples.

Operationally, atoms of background knowledge predicates become literals in the clauses of the learned hypothesis.

The next lines in anbn.pl, after the background knowledge declarations, generate positive and negative examples for 'S'/2. The positive examples are three strings in the a^nb^n language. These suffice for Thelma to learn a grammar for the entire language. The negative examples generator generates the empty list- no negative examples are needed.

The last two lines are the definitions of the two predicates defined as background knowledge, 'A'/2 and 'B'/2.

The need for inductive bias in machine learning

Background knowledge, metarules and order constraints impose a strong bias on the hypotheses that can be learned. The need for inductive bias in machine learning merits a longer discussion that is well beyond the scope of this introduction to Thelma. Suffice it to say that, without bias, there can be no learning. For a discussion, see (Mitchell, 1997). There are no machine learning algorithms that do not encode bias in some form- neural network architectures, Support Vector Machines' kernels, distance functions, Bayesian priors- are all instances of such bias encoding. The advantage of MIL and ILP algorithms in general is that inductive bias has a clear and precise definition that is easy to inspect and modify.

Future work

Thelma is still new and feature-light, for the time being.

The plan is to keep Thelma's interface and source code as straightforward and user-friendly as possible, sacrificing efficiency where necessary. That would make it more appropriate as an introductory system, to help teach interested beginners the basic concepts of MIL. For applications that require efficiency and speed, Metagol should be preferred.

Planned additional features include: an "experiment harness", comprising libraries for sampling, training, evaluation and logging; a MIL dataset generator; implementations of Plotkin's clause and program reduction algorithms; a library for bagging aggregation for learning in noisy domains; etc assorted libraries.

Further reading

A complete manual for Thelma with full usage instructions and more examples will be added at some point. Until then, consult the online documentation initiated when the load.pl file is loaded into Swi-Prolog.

For further reading on Meta-Interpretive Learning, see the reference section.

Bibliography and references

1. S.H. Muggleton, D. Lin, N. Pahlavi, and A. Tamaddoni-Nezhad. Meta-interpretive learning: application to grammatical inference. Machine Learning, 94:25-49, 2014

2. S.H. Muggleton, D. Lin, and A. Tamaddoni-Nezhad. Meta-interpretive learning of higher-order dyadic datalog: Predicate invention revisited. Machine Learning, 100(1):49-73, 2015

3. Metagol System Andrew Cropper and Stephen Muggleton, 2016

4. Knuth, D., & Bendix, P. (1970). Simple word problems in universal algebras. In J. Leech (Ed.), Computational problems in abstract algebra (pp. 263–297). Oxford: Pergamon.

5. Machine Learning, Tom Mitchell, McGraw Hill, 1997.

6. Zhang, T., Sipma, H., & Manna, Z. (2005). The decidability of the first-order theory of Knuth–Bendix order. In Automated deduction-CADE-20 (pp. 738–738). Berlin: Springer.